Qwen3-Coder-Next
(qwen.ai)604 points by danielhanchen 13 hours ago
604 points by danielhanchen 13 hours ago
We need a new word, not "local model" but "my own computers model" CapEx based
This distinction is important because some "we support local model" tools have things like ollama orchestration or use the llama.cpp libraries to connect to models on the same physical machine.
That's not my definition of local. Mine is "local network". so call it the "LAN model" until we come up with something better. "Self-host" exists but this usually means more "open-weights" as opposed to clamping the performance of the model.
It should be defined as ~sub-$10k, using Steve Jobs megapenny unit.
Essentially classify things as how many megapennies of spend a machine is that won't OOM on it.
That's what I mean when I say local: running inference for 'free' somewhere on hardware I control that's at most single digit thousands of dollars. And if I was feeling fancy, could potentially fine-tune on the days scale.
A modern 5090 build-out with a threadripper, nvme, 256GB RAM, this will run you about 10k +/- 1k. The MLX route is about $6000 out the door after tax (m3-ultra 60 core with 256GB).
Lastly it's not just "number of parameters". Not all 32B Q4_K_M models load at the same rate or use the same amount of memory. The internal architecture matters and the active parameter count + quantization is becoming a poorer approximation given the SOTA innovations.
What might be needed is some standardized eval benchmark against standardized hardware classes with basic real world tasks like toolcalling, code generation, and document procesing. There's plenty of "good enough" models out there for a large category of every day tasks, now I want to find out what runs the best
Take a gen6 thinkpad P14s/macbook pro and a 5090/mac studio, run the benchmark and then we can say something like "time-to-first-token/token-per-second/memory-used/total-time-of-test" and rate this as independent from how accurate the model was.
You can run plenty of models on a $10K machine or even a lot less than that, it all depends how much you want to wait for results. Streaming weights from SSD storage using mmap() is already a reality when running the largest and sparsest models. You can save even more on memory by limiting KV caching at the cost of extra compute, and there may be ways to push RAM savings even higher simply by tweaking the extent to which model activations are recomputed as needed.
Yeah there's a lot of people that advocate for really slow inference on cheap infra. That's something else that should be expressed in this fidelity
Because honestly I don't care about 0.2 tps for my use cases although I've spoken with many who are fine with numbers like that.
At least the people I've talked to they talk about how if they have a very high confidence score that the model will succeed they don't mind the wait.
Essentially a task failure is 1 in 10, I want to monitor and retry.
If it's 1 in 1000, then I can walk away.
The reality is most people don't have a bearing on what this order of magnitude actually is for a given task. So unless you have high confidence in your confidence score, slow is useless
But sometimes you do...
For context on what cloud API costs look like when running coding agents:
With Claude Sonnet at $3/$15 per 1M tokens, a typical agent loop with ~2K input tokens and ~500 output per call, 5 LLM calls per task, and 20% retry overhead (common with tool use): you're looking at roughly $0.05-0.10 per agent task.
At 1K tasks/day that's ~$1.5K-3K/month in API spend.
The retry overhead is where the real costs hide. Most cost comparisons assume perfect execution, but tool-calling agents fail parsing, need validation retries, etc. I've seen retry rates push effective costs 40-60% above baseline projections.
Local models trading 50x slower inference for $0 marginal cost start looking very attractive for high-volume, latency-tolerant workloads.
On the other hand, Deepseek V3.2 is $0.38 per million tokens output. And on openrouter, most providers serve it at 20 tokens/sec.
At 20t/s over 1 month, that's... $19something running literally 24/7. In reality it'd be cheaper than that.
I bet you'd burn more than $20 in electricity with a beefy machine that can run Deepseek.
The economics of batch>1 inference does not go in favor of consumers.
I don't even need "open weights" to run on hardware I own.
I am fine renting an H100 (or whatever), as long as I theoretically have access to and own everything running.
I do not want my career to become dependent upon Anthropic.
Honestly, the best thing for "open" might be for us to build open pipes and services and models where we can rent cloud. Large models will outpace small models: LLMs, video models, "world" models, etc.
I'd even be fine time-sharing a running instance of a large model in a large cloud. As long as all the constituent pieces are open where I could (in theory) distill it, run it myself, spin up my own copy, etc.
I do not deny that big models are superior. But I worry about the power the large hyperscalers are getting while we focus on small "open" models that really can't match the big ones.
We should focus on competing with large models, not artisanal homebrew stuff that is irrelevant.
> I do not want my career to become dependent upon Anthropic
As someone who switches between Anthropic and ChatGPT depending on the month and has dabbled with other providers and some local LLMs, I think this fear is unfounded.
It's really easy to switch between models. The different models have some differences that you notice over time but the techniques you learn in one place aren't going to lock you into a provider anywhere.
OOM is a pretty terrible benchmark too, though. You can build a DDR4 machine that "technically" loads 256gb models for maybe $1000 used, but then you've got to account for the compute aspect and that's constrained by a number of different variables. A super-sparse model might run great on that DDR4 machine, whereas a 32b model would cause it to chug.
There's just not a good way to visualize the compute needed, with all the nuance that exists. I think that trying to create these abstractions are what leads to people impulse buying resource-constrained hardware and getting frustrated. The autoscalers have a huge advantage in this field that homelabbers will never be able to match.
> time-to-first-token/token-per-second/memory-used/total-time-of-test
Would it not help with the DDR4 example though if we had more "real world" tests?
Maybe, but even that fourth-order metric is missing key performance details like context length and model size/sparsity.
The bigger takeaway (IMO) is that there will never really be hardware that scales like Claude or ChatGPT does. I love local AI, but it stresses the fundamental limits of on-device compute.
I run Qwen3-Coder-30B-A3B-Instruct gguf on a VM with 13gb RAM and a 6gb RTX 2060 mobile GPU passed through to it with ik_llama, and I would describe it as usable, at least. It's running on an old (5 years, maybe more) Razer Blade laptop that has a broken display and 16gb RAM.
I use opencode and have done a few toy projects and little changes in small repositories and can get pretty speedy and stable experience up to a 64k context.
It would probably fall apart if I wanted to use it on larger projects, but I've often set tasks running on it, stepped away for an hour, and had a solution when I return. It's definitely useful for smaller project, scaffolding, basic bug fixes, extra UI tweaks etc.
I don't think "usable" a binary thing though. I know you write lot about this, but it'd be interesting to understand what you're asking the local models to do, and what is it about what they do that you consider unusable on a relative monster of a laptop?
I've had usable results with qwen3:30b, for what I was doing. There's definitely a knack to breaking the problem down enough for it.
What's interesting to me about this model is how good it allegedly is with no thinking mode. That's my main complaint about qwen3:30b, how verbose its reasoning is. For the size it's astonishing otherwise.
Honestly I've been completely spoiled by Claude Code and Codex CLI against hosted models.
I'm hoping for an experience where I can tell my computer to do a thing - write a code, check for logged errors, find something in a bunch of files - and I get an answer a few moments later.
Setting a task and then coming back to see if it worked an hour later is too much friction for me!
> I still haven't experienced a local model that fits on my 64GB MacBook Pro and can run a coding agent like Codex CLI or Claude code well enough to be useful
I've had mild success with GPT-OSS-120b (MXFP4, ends up taking ~66GB of VRAM for me with llama.cpp) and Codex.
I'm wondering if maybe one could crowdsource chat logs for GPT-OSS-120b running with Codex, then seed another post-training run to fine-tune the 20b variant with the good runs from 120b, if that'd make a big difference. Both models with the reasoning_effort set to high are actually quite good compared to other downloadable models, although the 120b is just about out of reach for 64GB so getting the 20b better for specific use cases seems like it'd be useful.
This is a common problem for people trying to run the GPT-oss models themselves. Reposting my comment here:
GPT-oss-120B was also completely failing for me, until someone on reddit pointed out that you need to pass back in the reasoning tokens when generating a response. One way to do this is described here:
https://openrouter.ai/docs/guides/best-practices/reasoning-t...
Once I did that it started functioning extremely well, and it's the main model I use for my homemade agents.
Many LLM libraries/services/frontends don't pass these reasoning tokens back to the model correctly, which is why people complain about this model so much. It also highlights the importance of rolling these things yourself and understanding what's going on under the hood, because there's so many broken implementations floating around.
You are describing distillation, there are better ways to do it, and it was done in the past, Deepseek distilled onto Qwen.
Yeah, Apple hardware don't seem ideal for LLMs that are large, give it a go with a dedicated GPU if you're inclined and you'll see a big difference :)
What are some good GPUs to look for if you're getting started?
I configured Claude Code to use a local model (ollama run glm-4.7-flash) that runs really well on a 32G M2Pro macmini. Maybe my standards are too low, but I was using that combination to clean up the code, make improvements, and add docs and tests to a bunch of old git repo experiment projects.
Did you have to do anything special to get it to work? I tried and it would just bug out, things like respond with JSON strings summarizing what I asked of it or just outright getting things wrong entirely. For example, I asked it to summarize what a specific .js file did and it provided me with new code it made up based on the file name...
I wonder if the future in ~5 years is almost all local models? High-end computers and GPUs can already do it for decent models, but not sota models. 5 years is enough time to ramp up memory production, consumers to level-up their hardware, and models to optimize down to lower-end hardware while still being really good.
Opensource or local models will always heavily lag frontier.
Who pays for a free model? GPU training isn't free!
I remember early on people saying 100B+ models will run on your phone like nowish. They were completely wrong and I don't think it's going to ever really change.
People always will want the fastest, best, easiest setup method.
"Good enough" massively changes when your marketing team is managing k8s clusters with frontier systems in the near future.
I don't think this is as true as you think.
People do not care about the fastest and best past a point.
Let's use transportation as an analogy. If all you have is a horse, a car is a massive improvement. And when cars were just invented, a car with a 40mph top speed was a massive improvement over one with a 20mph top speed and everyone swapped.
While cars with 200mph top speeds exist, most people don't buy them. We all collectively decided that for most of us, most of the time, a top speed of 110-120 was plenty, and that envelope stopped being pushed for consumer vehicles.
If what currently takes Claude Opus 10 minutes to do can be done is 30ms, then making something that can do it in 20ms isn't going to be enough to get everyone to pay a bunch of extra money for.
Companies will buy the cheapest thing that meets their needs. SOTA models right now are much better than the previous generation but we have been seeing diminishing returns in the jump sizes with each of the last couple generations. If the gap between current and last gen shrinks enough, then people won't pay extra for current gen if they don't need it. Just like right now you might use Sonnet or Haiku if you don't think you need Opus.
This is the assumption of a hard plateu we can effectively optimize forever towards while possible we havn't seen it.
Again my point is "good enough" changes as possibilities open. Marketing teams running entire infra stacks is an insane idea today but may not be in the future.
You could easily code with a local model similar to gpt 4 or 3 now but I will 10-100x your performance with a frontier model and that will fundamentally not change.
Hmmm but maybe there's an argument of a static task. Once a model hits that ability of that specific task you can optimize it into a smaller model. So I guess I buy the argument for people working on statically capped conplexity tasks?
PII detection for example, a <500M model will outperform a 1-8B param model on that narrow task. But at the same time just a pii detection bot is not a product anymore. So yes a opensource one does it but as a result its fundamentally less valuable and I need to build higher and larger products for the value?
Gpt3.5 as used in the first commercially available chat gpt is believed to be hundreds of billions of parameters. There are now models I can run on my phone that feel like they have similar levels of capability.
Phones are never going to run the largest models locally because they just don't have the size, but we're seeing improvements in capability at small sizes over time that mean that you can run a model on your phone now that would have required hundreds of billions of parameters less than 6 years ago.
I don't know about frontier, I code nowadays a lot using Opus 4.5, in a way that I instruct it to do something (like complex refactor etc) - I like that it's really good at actually doing what its told and only occasionally do I have to fight it when it goes off the rails. It also does not hallucinate all that much in my experience (Im writing Js, YMMV with other languages), and is good at spotting dumb mistakes.
That said, I'm not sure if this capability is only achievable in huge frontier models, I would be perfectly content using a model that can do this (acting as a force multiplier), and not much else.
I think we'll eventually find a way to make the cycle smaller, so instead of writing a stackoverflow post in 2024 and using a model trained on it in 2025 I'll be contributing to the expertise of a distributed-model-ish-thing on Monday and benefitting from that contribution on Tuesday.
When that happens, the most powerful AI will be whichever has the most virtuous cycles going with as wide a set of active users as possible. Free will be hard to compete with because raising the price will exclude the users that make it work.
Until then though, I think you're right that open will lag.
I'm hoping so. What's amazing is that with local models you don't suffer from what I call "usage anxiety" where I find myself saving my Claude usage for hypothetical more important things that may come up, or constantly adjusting prompts and doing some manual work myself to spare token usage.
Having this power locally means you can play around and experiment more without worries, it sounds like a wonderful future.
Plus a long queue of yet-undiscovered architectural improvements
I'm suprised there isn't more "hope" in this area. Even things like the GPT Pro models; surely that sort of reasoning/synthesis will eventually make its way into local models. And that's something that's already been discovered.
Just the other day I was reading a paper about ANNs whose connections aren't strictly feedforward but, rather, circular connections proliferate. It increases expressiveness at the (huge) cost of eliminating the current gradient descent algorithms. As compute gets cheaper and cheaper, these things will become feasible (greater expressiveness, after all, equates to greater intelligence).
A lot of manufacturers are bailing on consumer lines to focus on enterprise from what I've read. Not great.
Even without leveling up hardware, 5 years is a loooong time to squeeze the juice out of lower-end model capability. Although in this specific niche we do seem to be leaning on Qwen a lot.
Why don't you try it out in Opencode? It's possible to hook up the openrouter api, and some providers have started to host it there [1]. It's not yet available in opencode's model list [2].
Opencode's /connect command has a big list of providers, openrouter is on there.
I can't get Codex CLI or Claude Code to use small local models and to use tools. This is because those tools use XML and the small local models have JSON tool use baked into them. No amount of prompting can fix it.
In a day or two I'll release my answer to this problem. But, I'm curious, have you had a different experience where tool use works in one of these CLIs with a small local model?
You mean Qwen3-Coder-Next? I haven't tried that model itself, yet, because I assume it's too big for me. I have a modest 16GB MacBook Air so I'm restricted to really small stuff. I'm thinking about buying a machine with a GPU to run some of these.
Anywayz, maybe I should try some other models. The ones that haven't worked for tool calling, for me are:
Llama3.1
Llama3.2
Qwen2.5-coder
Qwen3-coder
All these in 7b, 8b, or sometimes 30b (painfully) models.
I should also note that I'm typically using Ollama. Maybe LM Studio or llama.cpp somehow improve on this?
Surely the answer is a very small proxy server between the two?
They run fairly well for me on my 128GB Framework Desktop.
what do you run this on if I may ask? lmstudio, ollama, lama? which cli?
Unfortunately Qwen3-next is not well supported on Apple silicon, it seems the Qwen team doesn't really care about Apple.
On M1 64GB Q4KM on llama.cpp gives only 20Tok/s while on MLX it is more than twice as fast. However, MLX has problems with kv cache consistency and especially with branching. So while in theory it is twice as fast as llama.cpp it often does the PP all over again which completely trashes performance especially with agentic coding.
So the agony is to decide whether to endure half the possible speed but getting much better kv-caching in return. Or to have twice the speed but then often you have again to sit through prompt processing.
But who knows, maybe Qwen gives them a hand? (hint,hint)
KV caching means that when you have 10k prompt, all follow up questions return immediately - this is standard with all inference engines.
Now if you are not happy with the last answer, you maybe want to simply regenerate it or change your last question - this is branching of the conversation. Llama.cpp is capable of re-using the KV cache up to that point while MLX does not (I am using MLX server from MLX community project). I haven't tried with LMStudio. Maybe worth a try, thanks for the heads-up.
It is the buffer implementation. [u1 10kTok]->[a1]->[u2]->[a2]. If you branch between the assistant1 and user2 answers then MLX does reprocess the u1 prompt of let's say 10k tokens while llama.cpp does not.
I just tested with GGUF and MLX of Qwen3-Coder-Next with llama.cpp and now with LMStudio. As I do branching very often, it is highly annoying for me to the point of being unusable. Q3-30B is much more usable then on Mac - but by far not as powerful.
There’s this issue/outstanding PR: https://github.com/lmstudio-ai/mlx-engine/pull/188#issuecomm...
It works reasonably well for general tasks, so we're definitely getting there! Probably Qwen3 CLI might be better suited, but haven't tested it yet.
I don't buy this. I've long wondered if the larger models, while exhibiting more useful knowledge, are not more wasteful as we greedily explore the frontier of "bigger is getting us better results, make it bigger". Qwen3-Coder-Next seems to be a point for that thought: we need to spend some time exploring what smaller models are capable of.
Perhaps I'm grossly wrong -- I guess time will tell.
You are not wrong, small models can be trained for niche use cases and there are lots of people and companies doing that. The problem is that you need one of those for each use case whereas the bigger models can cover a bigger problem space.
There is also the counter-intuitive phenomenon where training a model on a wider variety of content than apparently necessary for the task makes it better somehow. For example, models trained only on English content exhibit measurably worse performance at writing sensible English than those trained on a handful of languages, even when controlling for the size of the training set. It doesn't make sense to me, but it probably does to credentialed AI researchers who know what's going on under the hood.
There is (must be - information theory) a size/capacity efficiency frontier. There is no particular reason to think we're anywhere near it right now.
For those interested, made some Dynamic Unsloth GGUFs for local deployment at https://huggingface.co/unsloth/Qwen3-Coder-Next-GGUF and made a guide on using Claude Code / Codex locally: https://unsloth.ai/docs/models/qwen3-coder-next
Nice! Getting ~39 tok/s @ ~60% GPU util. (~170W out of 303W per nvtop).
System info:
$ ./llama-server --version
ggml_vulkan: Found 1 Vulkan devices:
ggml_vulkan: 0 = Radeon RX 7900 XTX (RADV NAVI31) (radv) | uma: 0 | fp16: 1 | bf16: 0 | warp size: 64 | shared memory: 65536 | int dot: 1 | matrix cores: KHR_coopmat
version: 7897 (3dd95914d)
built with GNU 11.4.0 for Linux x86_64
$ ./llama-server --host 0.0.0.0 --port 2000 --no-warmup \
-hf unsloth/Qwen3-Coder-Next-GGUF:UD-Q4_K_XL \
--jinja --temp 1.0 --top-p 0.95 --min-p 0.01 --top-k 40 --fit on \
--ctx-size 32768Super cool! Also with `--fit on` you don't need `--ctx-size 32768` technically anymore - llama-server will auto determine the max context size!
What am I missing here? I thought this model needs 46GB of unified memory for 4-bit quant. Radeon RX 7900 XTX has 24GB of memory right? Hoping to get some insight, thanks in advance!
MoEs can be efficiently split between dense weights (attention/KV/etc) and sparse (MoE) weights. By running the dense weights on the GPU and offloading the sparse weights to slower CPU RAM, you can still get surprisingly decent performance out of a lot of MoEs.
Not as good as running the entire thing on the GPU, of course.
Hi Daniel, I've been using some of your models on my Framework Desktop at home. Thanks for all that you do.
Asking from a place of pure ignorance here, because I don't see the answer on HF or in your docs: Why would I (or anyone) want to run this instead of Qwen3's own GGUFs?
Thanks! Oh Qwen3's own GGUFs also works, but ours are dynamically quantized and calibrated with a reasonably large diverse dataset, whilst Qwen's ones are not - see https://unsloth.ai/docs/basics/unsloth-dynamic-2.0-ggufs
What is the difference between the UD and non-UD files?
UD stands for "Unsloth-Dynamic" which upcasts important layers to higher bits. Non UD is just standard llama.cpp quants. Both still use our calibration dataset.
Please consider authoring a single, straightforward introductory-level page somewhere that explains what all the filename components mean, and who should use which variants.
The green/yellow/red indicators for different levels of hardware support are really helpful, but far from enough IMO.
What is your definition of "important" in this context?
Oh we wrote about it here: https://unsloth.ai/docs/basics/unsloth-dynamic-2.0-ggufs
Good results with your Q8_0 version on 96GB RTX 6000 Blackwell. It one-shotted the Flappy Bird game and also wrote a good Wordle clone in four shots, all at over 60 tps. Thanks!
Is your Q8_0 file the same as the one hosted directly on the Qwen GGUF page?
Nice! Yes Q8_0 is similar - the others are different since they use a calibration dataset.
I got this running locally using llama.cpp from Homebrew and the Unsloth quantized model like this:
brew upgrade llama.cpp # or brew install if you don't have it yet
llama-cli \
-hf unsloth/Qwen3-Coder-Next-GGUF:UD-Q4_K_XL \
--fit on \
--seed 3407 \
--temp 1.0 \
--top-p 0.95 \
--min-p 0.01 \
--top-k 40 \
--jinja
llama-server \
-hf unsloth/Qwen3-Coder-Next-GGUF:UD-Q4_K_XL \
--fit on \
--seed 3407 \
--temp 1.0 \
--top-p 0.95 \
--min-p 0.01 \
--top-k 40 \
--jinja
Yes sadly that sometimes happens - the issue is Codex CLI / Claude Code were designed for GPT / Claude models specifically, so it'll be hard for OSS models directly to utilize the full spec / tools etc, and might get loops sometimes - I would maybe try the MXFP4_MOE quant to see if it helps, and maybe try Qwen CLI (was planning to make a guide for it as well)
I guess until we see the day OSS models truly utilize Codex / CC very well, then local models will really take off
It’s hard to elaborate just how wild this model might be if it performs as claimed. The claims are this can perform close to Sonnet 4.5 for assisted coding (SWE bench) while using only 3B active parameters. This is obscenely small for the claimed performance.
I experimented with the Q2 and Q4 quants. First impression is that it's amazing we can run this locally, but it's definitely not at Sonnet 4.5 level at all.
Even for my usual toy coding problems it would get simple things wrong and require some poking to get to it.
A few times it got stuck in thinking loops and I had to cancel prompts.
This was using the recommended settings from the unsloth repository. It's always possible that there are some bugs in early implementations that need to be fixed later, but so far I don't see any reason to believe this is actually a Sonnet 4.5 level model.
Wonder where it falls on the Sonnet 3.7/4.0/4.5 continuum.
3.7 was not all that great. 4 was decent for specific things, especially self contained stuff like tests, but couldn't do a good job with more complex work. 4.5 is now excellent at many things.
If it's around the perf of 3.7, that's interesting but not amazing. If it's around 4, that's useful.
It feels more like Haiku level than Sonnet 4.5 from my playing with it.
Should be possible with optimised models, just drop all "generic" stuff and focus on coding performance.
There's no reason for a coding model to contain all of ao3 and wikipedia =)
There is: It works (even if we can't explain why right now).
If we knew how to create a SOTA coding model by just putting coding stuff in there, that is how we would build SOTA coding models.
I think I like coding models that know a lot about the world. They can disambiguate my requirements and build better products.
That's what Meta thought initially too, training codellama and chat llama separately, and then they realized they're idiots and that adding the other half of data vastly improves both models. As long as it's quality data, more of it doesn't do harm.
Besides, programming is far from just knowing how to autocomplete syntax, you need a model that's proficient in the fields that the automation is placed in, otherwise they'll be no help in actually automating it.
But as far as I know, that was way before tool calling was a thing.
I'm more bullish about small and medium sized models + efficient tool calling than I'm about LLMs too large to be run at home without $20k of hardware.
The model doesn't need to have the full knowledge of everything built into it when it has the toolset to fetch, cache and read any information available.
But... but... I need my coding model to be able to write fanfiction in the comments...
There have been advances recently (last year) in scaling deep rl by a significant amount, their announcement is in line with a timeline of running enough experiments to figure out how to leverage that in post training.
Importantly, this isn’t just throwing more data at the problem in an unstructured way, afaik companies are getting as many got histories as they can and doing something along the lines of, get an llm to checkpoint pull requests, features etc and convert those into plausible input prompts, then run deep rl with something which passes the acceptance criteria / tests as the reward signal.
It literally always is. HN Thought DeepSeek and every version of Kimi would finally dethrone the bigger models from Anthropic, OpenAI, and Google. They're literally always wrong and average knowledge of LLMs here is shockingly low.
Nobody has been saying they'd be dethroned. We're saying they're often "good enough" for many use cases, and that they're doing a good job of stopping the Big Guys from creating a giant expensive moat around their businesses.
Chinese labs are acting as a disruption against Altman etcs attempt to create big tech monopolies, and that's why some of us cheer for them.
I got the Qwen3 Coder 30B running locally on mac Mac M4 Max 36GB. It was slow, but it worked and did do some decent stuff: https://www.youtube.com/watch?v=7mAPaRbsjTU
Video is speed up. I ran it through LM Studio and then OpenCode. Wrote a bit about how I set it all up here: https://www.tommyjepsen.com/blog/run-llm-locally-for-coding
3B active parameters, and slightly worse than GLM 4.7. On benchmarks. That's pretty amazing! With better orchestration tools being deployed, I've been wondering if faster, dumber coding agents paired with wise orchestrators might be overall faster than using the say opus 4.5 on the bottom for coding. At least we might want to deploy to these guys for simple tasks.
It's getting a lot easier to do this using sub-agents with tools in Claude. I have a fleet of Mastra agents (TypeScript). I use those agents inside my project as CLI tools to do repetitive tasks that gobble tokens such as scanning code, web search, library search, and even SourceGraph traversal.
Overall, it's allowed me to maintain more consistent workflows as I'm less dependent on Opus. Now that Mastra has introduced the concept of Workspaces, which allow for more agentic development, this approach has become even more powerful.
> just (expensive) magic trick
Related: as an actual magician, although no longer performing professionally, I was telling another magician friend the other day that IMHO, LLMs are the single greatest magic trick ever invented judging by pure deceptive power. Two reasons:
1. Great magic tricks exploit flaws in human perception and reasoning by seeming to be something they aren't. The best leverage more than one. By their nature, LLMs perfectly exploit the ways humans assess intelligence in themselves and others - knowledge recall, verbal agility, pattern recognition, confident articulation, etc. No other magic trick stacks so many parallel exploits at once.
2. But even the greatest magic tricks don't fool their inventors. David Copperfield doesn't suspect the lady may be floating by magic. Yet, some AI researchers believe the largest, most complex LLMs actually demonstrate emergent thinking and even consciousness. It's so deceptive it even fools people who know how it works. To me, that's a great fucking trick.
Time will tell. All this stuff will get more adoption when Anthropic, Google and OpenAI raise prices.
They can only raise prices as long as people buy their subscriptions / pay for their api. The Chinese labs are closing in on the SOTA models (I would say they are already there) and offer insane cheap prices for their subscriptions. Vote with your wallet.
What is the best place to see local model rankings? The benchmarks seem so heavily gamed that I am willing to believe the “objective” rankings are a lie and personal reviews are more meaningful.
Are there any clear winners per domain? Code, voice-to-text, text-to-voice, text editing, image generation, text summarization, business-text-generation, music synthesis, whatever.
17t/s on a laptop with 6GB VRAM and DDR5 system memory. Maximum of 100k context window (then it saturates VRAM). Quite amazing, but tbh I'll still use inference providers, because it's too slow and it's my only machine with "good" specs :)
cat docker-compose.yml
services:
llamacpp:
volumes:
- llamacpp:/root
container_name: llamacpp
restart: unless-stopped
image: ghcr.io/ggml-org/llama.cpp:server-cuda
network_mode: host
command: |
-hf unsloth/Qwen3-Coder-Next-GGUF:Q4_K_XL --jinja --cpu-moe --n-gpu-layers 999 --ctx-size 102400 --temp 1.0 --top-p 0.95 --min-p 0.01 --top-k 40 --fit on
# unsloth/gpt-oss-120b-GGUF:Q2_K
deploy:
resources:
reservations:
devices:
- driver: nvidia
count: all
capabilities: [gpu]
volumes:
llamacpp:Using lmstudio-community/Qwen3-Coder-Next-GGUF:Q8_0 I'm getting up to 32 tokens/s on Strix Halo, with room for 128k of context (out of 256k that the model can manage).
From very limited testing, it seems to be slightly worse than MiniMax M2.1 Q6 (a model about twice its size). I'm impressed.
Works great for these type of MOE models. The ability to have large amounts of VRAM let you run different models in parallel easily, or to have actually useful context sizes. Dense models can get sluggish though. AMD's ROCm support has been a little rough for Stable Diffusion stuff (memory issues leading to application stability problems) but it's worked well with LLMs, as does Vulkan.
I wish AMD would get around to adding NPU support in Linux for it though, it has more potential that could be unlocked.
I'm getting similar numbers on NVIDIA Spark around 25-30 tokens/sec output, 251 token/sec prompt processing... but I'm running with the Q4_K_XL quant. I'll try the Q8 next, but that would leave less room for context.
I tried FP8 in vLLM and it used 110GB and then my machine started to swap when I hit it with a query. Only room for 16k context.
I suspect there will be some optimizations over the next few weeks that will pick up the performance on these type of machines.
I have it writing some Rust code and it's definitely slower than using a hosted model but it's actually seeming pretty competent. These are the first results I've had on a locally hosted model that I could see myself actually using, though only once the speed picks up a bit.
I suspect the API providers will offer this model for nice and cheap, too.
llama.cpp is giving me ~35tok/sec with the unsloth quants (UD-Q4_K_XL, elsewhere in this thread) on my Spark. FWIW my understanding and experience is that llama.cpp seems to give slight better performance for "single user" workloads, but I'm not sure why.
I'm asking it to do some analysis/explain some Rust code in a rather large open source project and it's working nicely. I agree this is a model I could possibly, maybe use locally...
Yeah I got 35-39tok/sec for one shot prompts, but for real-world longer context interactions through opencode it seems to be averaging out to 20-30tok/sec. I tried both MXFP4 and Q4_K_XL, no big difference, unfortunately.
--no-mmap --fa on options seemed to help, but not dramatically.
As with everything Spark, memory bandwidth is the limitation.
I'd like to be impressed with 30tok/sec but it's sort of a "leave it overnight and come back to the results" kind of experience, wouldn't replace my normal agent use.
However I suspect in a few days/weeks DeepInfra.com and others will have this model (maybe Groq, too?), and will serve it faster and for fairly cheap.
This is model 12188, which claims to rival SOTA models while not even being in the same league.
In terms of intelligence per compute, it’s probably the best model I can realistically run locally on my laptop for coding. It’s solid for scripting and small projects.
I tried it on a mid-size codebase (~50k LOC), and the context window filled up almost immediately, making it basically unusable unless you’re extremely explicit about which files to touch. I tested it with a 8k context window but will try again with 32k and see if it becomes more practical.
I think the main blocker for using local coding models more is the context window. A lot of work is going into making small models “smarter,” but for agentic coding that only gets you so far. No matter how smart the model is, an agent will blow through the context as soon as it reads a handful of files.
Benchmarks using DGX Spark on vLLM 0.15.1.dev0+gf17644344
FP8: https://huggingface.co/Qwen/Qwen3-Coder-Next-FP8
Sequential (single request)
Prompt Gen Prompt Processing Token Gen
Tokens Tokens (tokens/sec) (tokens/sec)
------ ------ ----------------- -----------
521 49 3,157 44.2
1,033 83 3,917 43.7
2,057 77 3,937 43.6
4,105 77 4,453 43.2
8,201 77 4,710 42.2
Parallel (concurrent requests)
pp4096+tg128 (4K context, 128 gen):
n t/s
-- ----
1 28.5
2 39.0
4 50.4
8 57.5
16 61.4
32 62.0
pp8192+tg128 (8K context, 128 gen):
n t/s
-- ----
1 21.6
2 27.1
4 31.9
8 32.7
16 33.7
32 31.7I tried the FP8 in vLLM on my Spark and although it fit in memory, I started swapping once I actually tried to run any queries, and, yeah, could not have a context larger than 8k.
I figured out later this is because vLLM apparently de-quantizes to BF16 at runtime, so pointless to run the FP8?
I get about 30-35 tok/second using llama.cpp and a 4-bit quant. And a 200+k context, using only 50GB of RAM.
Running llama.cpp rather than vLLM, it's happy enough to run the FP8 variant with 200k+ context using about 90GB vram
yeah, what did you get for tok/sec there though? Memory bandwidth is the limitation with these devices. With 4 bit I didn't get over 35-39 tok/sec, and averaged more like 30 when doing actual tool use with opencode. I can't imagine fp8 being faster.
As always, the Qwen team is pushing out fantastic content
Hope they update the model page soon https://chat.qwen.ai/settings/model
> "content"
Sorry, but we're talking about models as content now? There's almost always a better word than "content" if you're describing something that's in tech or online.
I kind of lost interest in local models. Then Anthropic started saying I’m not allowed to use my Claude Code subscription with my preferred tools and it reminded me why we need to support open tools and models. I’ve cancelled my CC subscription, I’m not paying to support anticompetitive behaviour.
> Then Anthropic started saying I’m not allowed to use my Claude Code subscription with my preferred tools
To be clear, since this confuses a lot of people in every thread: Anthropic will let you use their API with any coding tools you want. You just have to go through the public API and pay the same rate as everyone else. They have not "blocked" or "banned" any coding tools from using their API, even though a lot of the clickbait headlines have tried to insinuate as much.
Anthropic never sold subscription plans as being usable with anything other than their own tools. They were specifically offered as a way to use their own apps for a flat monthly fee.
They obviously set the limits and pricing according to typical use patterns of these tools, because the typical users aren't maxing out their credits in every usage window.
Some of the open source tools reverse engineered the protocol (which wasn't hard) and people started using the plans with other tools. This situation went on for a while without enforcement until it got too big to ignore, and they began protecting the private endpoints explicitly.
The subscription plans were never sold as a way to use the API with other programs, but I think they let it slide for a while because it was only a small number of people doing it. Once the tools started getting more popular they started closing loopholes to use the private API with other tools, which shouldn't really come as a surprise.
The anticompetitive move would be not running their software if ‘which codex’ evaluated to showing a binary and then not allow you to use it due to its presence. Companies are allowed to set pricing and not let you borrow the jet to fly to a not approved destination. This distortion is just wrong as a premise. They are being competitive by making a superior tool and their business model is “no one else sells Claude” and they are pretty right to do this IMO.
Anticompetitive behavior has been normalized in our industry, doesn't make it not anticompetitive. It's a restriction that's meant to make it harder to compete with other parts of their offering. The non-anticompetitive approach would be to offer their subscription plans with a certain number of tokens every month, and then make Claude Code the most efficient with the tokens, to let it compete on its own merits.
> Anthropic will let you use their API with any coding tools you want
No, in 2026, even with their API plan the create key is disabled for most orgs, you basically have to ask your admin to give you a key to use something other than Claude Code. You can imagine how that would be a problem.
That’s not an Anthropic problem, that’s a problem with whomever you work for.
Yes, exactly. The discourse has been so far off the rails now.
from what i remember, i couldnt actually use claude code with the subscription when i subscribed. i could only use it with third party tools.
eventually they added subscription support and that worked better than cline or kilo, but im still not clear what anthropic tools the subscription was actually useful for
The question I pose is this: if they're willing to start building walls this early in the game while they've still got plenty of viable competitors, and are at most 6 months ahead, how will they treat us if they achieve market dominance?
Some people think LLMs are the final frontier. If we just give in and let Anthropic dictate the terms to us we're going to experience unprecedented enshittification. The software freedom fight is more important than ever. My machine is sovereign; Anthropic provides the API, everything I do on my machine is my concern.
I don't get why so much mental gymnastics is done to avoid the fact that locking their lower prices to effectively subsidize their shitty product is the anti competitive behavior.
They simply don't want to compete, they want to force the majority of people that can't spend a lot on tokens to use their inferior product.
Why build a better product if you control the cost?
You gave up some convenience to avoid voting for a bad practice with your wallet. I admire this, try to consistently do this when reasonably feasible.
Problem is, most people don't do this, choosing convenience at any given moment without thinking about longer-term impact. This hurts us collectively by letting governments/companies, etc tighten their grip over time. This comes from my lived experience.
As I get older I more and more view convenience as the enemy of good. Luckily (or unluckily for some) a lot of the tradeoffs we are asked to make in the name of convenience are increasingly absurd. I have an easier and easier time going without these Faustian bargains.
IMHO The question is: who is in control? The user, or the profit-seeking company/control-seeking government? There is nothing we can do to prevent companies from seeking profit. What we can do is to prefer tools that we control, if that choice is not available, then tools that we can abandon when we want, over tools that remove our control AND abandoning them would be prohibitively difficult.
Claude Opus 4.5 by far is the most capable development model. I've been using it mainly via Claude Code, and with Cursor.
I agree anticompetitive behavior is bad, but the productivity gains to be had by using Anthropic models and tools are undeniable.
Eventually the open tools and models will catch up, so I'm all for using them locally as well, especially if sensitive data or IP is involved.
I'd encourage you to try the -codex family with the highest reasoning.
I can't comment on Opus in CC because I've never bit the bullet and paid the subscription, but I have worked my way up to the $200/month Cursor subscription and the 5.2 codex models blow Opus out of the water in my experience (obviously very subjective).
I arrived at making plans with Opus and then implementing with the OpenAI model. The speed of Opus is much better for planning.
I'm willing to believe that CC/Opus is truly the overall best; I'm only commenting because you mentioned Cursor, where I'm fairly confident it's not. I'm basing my judgement on "how frequently does it do what I want the first time".
Thanks, I'll try those out. I've used Codex CLI itself on a few small projects as well, and fired it up on a feature branch where I had it implement the same feature that Claude Code did (they didn't see each other's implementations). For that specific case, the implementation Codex produced was simpler, and better for the immediate requirements. However, Claude's more abstracted solution may have held up better to changing requirements. Codex feels more reserved than Claude Code, which can be good or bad depending on the task.
I've tried nearly all the models, they all work best if and only if you will never handle the code ever again. They suck if you have a solution and want them to implement that solution.
I've tried explaining the implementation word and word and it still prefers to create a whole new implementation reimplementing some parts instead of just doing what I tell it to. The only time it works is if I actually give it the code but at that point there's no reason to use it.
There's nothing wrong with this approach if it actually had guarantees, but current models are an extremely bad fit for it.
> Claude Opus 4.5 by far is the most capable development model.
At the moment I have a personal Claude Max subscription and ChatGPT Enterprise for Codex at work. Using both, I feel pretty definitively that gpt-5.2-codex is strictly superior to Opus 4.5. When I use Opus 4.5 I’m still constantly dealing with it cutting corners, misinterpreting my intentions and stopping when it isn’t actually done. When I switched to Codex for work a few months ago all of those problems went away.
I got the personal subscription this month to try out Gas Town and see how Opus 4.5 does on various tasks, and there are definitely features of CC that I miss with Codex CLI (I can’t believe they still don’t have hooks), but I’ve cancelled the subscription and won’t renew it at the end of this month unless they drop a model that really brings them up to where gpt-5.2-codex is at.
I have literally the opposite experience and so does most of AI pilled twitter and the AI research community of top conferences (NeurIPS, ICLR, ICML, AAAI) Why does this FUD keep appearing on this site?
Edit: It's very true that the big 4 labs silently mess with their models and any action of that nature is extremely user hostile.
Probably because all of the major providers are constantly screwing around with their models, regardless of what they say.
It feels very close to a trade-off point.
I agree with all posts in the chain: Opus is good, Anthropic have burned good will, I would like to use other models...but Opus is too good.
What I find most frustrating is that I am not sure if it is even actual model quality that is the blocker with other models. Gemini just goes off the rails sometimes with strange bugs like writing random text continuously and burning output tokens, Grok seems to have system prompts that result in odd behaviour...no bugs just doing weird things, Gemini Flash models seem to output massive quantities of text for no reason...it is often feels like very stupid things.
Also, there are huge issues with adopting some of these open models in terms of IP. Third parties are running these models and you are just sending them all your code...with a code of conduct promise from OpenRouter?
I also don't think there needs to be a huge improvement in models. Opus feels somewhat close to the reasonable limit: useful, still outputs nonsense, misses things sometimes...there are open models that can reach the same 95th percentile but the median is just the model outputting complete nonsense and trying to wipe your file system.
The day for open models will come but it still feels so close and so far.
I do wonder if they locked things down due to people abusing their CC token.
I buy the theory that Claude Code is engineered to use things like token caching efficiently, and their Claude Max plans were designed with those optimizations in mind.
If people start using the Claude Max plans with other agent harnesses that don't use the same kinds of optimizations the economics may no longer have worked out.
(But I also buy that they're going for horizontal control of the stack here and banning other agent harnesses was a competitive move to support that.)
It should just burn quota faster then. Instead of blocking they should just mention that if you use other tools then your quota may reduce at 3x speed compared to cc. People would switch.
When I last checked a few months ago, Anthropic was the only provider that didn't have automatic prompt caching. You had to do it manually (and you could only set checkpoints a few times per context?), and most 3rd party stuff does not.
They seem to have started rejecting 3rd party usage of the sub a few weeks ago, before Claw blew up.
By the way, does anyone know about the Agents SDK? Apparently you can use it with an auth token, is anyone doing that? Or is it likely to get your account in trouble as well?
Absolutely. I installed clawdbot for just long enough to send a single message, and it burned through almost a quarter of my session allowance. That was enough for me. Meanwhile I can use CC comfortably for a few hours and I've only hit my token limit a few times.
I've had a similar experience with opencode, but I find that works better with my local models anyway.
I would be surprised if the primary reason for banning third party clients isn't because they are collecting training data via telemetry and analytics in CC. I know CC needlessly connects to google infrastructure, I assume for analytics.
If that was the real reason, why wouldn't they just make it so that if you don't correctly use caching you use up more of your limit?
In what way would it be abused? The usage limits apply all the same, they aren't client side, and hitting that limit is within the terms of the agreement with Anthropic.
The subscription services have assumptions baked in about the usage patterns; they're oversubscribed and subsidized. If 100% of subscriber customers use 100% of their tokens 100% of the time, their business model breaks. That's what wholesale / API tokens are for.
> hitting that limit is within the terms of the agreement with Anthropic
It's not, because the agreement says you can only use CC.
Taking umbrage as if it matters how I use the compute I'm paying for via the harness they want me to use it within as long as I'm just doing personal tasks I want to do for myself, not trying to power an apps API with it seems such a waste of their time to be focusing on and only causes brand perception damage with their customers.
Could have just turned a blind eye.
The loss of access shows the kind of power they'll have in the future. It's just a taste of what's to come.
If a company is going to automate our jobs, we shouldn't be giving them money and data to do so. They're using us to put ourselves out of work, and they're not giving us the keys.
I'm fine with non-local, open weights models. Not everything has to run on a local GPU, but it has to be something we can own.
I'd like a large, non-local Qwen3-Coder that I can launch in a RunPod or similar instance. I think on-demand non-local cloud compute can serve as a middle ground.
How do I "abuse" a token? I pass it to their API, the request executes, a response is returned, I get billed for it. That should be the end of the conversation.
(Edit due to rate-limiting: I see, thanks -- I wasn't aware there was more than one token type.)
You can buy this product, right here: https://platform.claude.com/docs/en/about-claude/pricing
That's not the product you buy when you a Claude Code token, though.
Claude Code supports using API credits, and you can turn on Extra Usage and use API credits automatically once your session limit is reached.
This confused me for a while, having two separate "products" which are sold differently, but can be used by the same tool.
Access is one of my concerns with coding agents - on the one hand I think they make coding much more accessible to people who aren't developers - on the other hand this access is managed by commercial entities and can be suspended for any reason.
I can also imagine a dysfunctional future where a developers spend half their time convincing their AI agents that the software they're writing is actually aligned with the model's set of values
Anthropic banned my account when I whipped up a solution to control Claude Code running on my Mac from my phone when I'm out and about. No commercial angle, just a tool I made for myself since they wouldn't ship this feature (and still haven't). I wasn't their biggest fanboy to begin with, but it gave me the kick in the butt needed to go and explore alternatives until local models get good enough that I don't need to use hosted models altogether.
I didn't like the existing SSH applications for iOS and I already have a local app that I made that I have open 24/7, so I added a screen that used xterm.js and Bun.spawn with Bun.Terminal to mirror the process running on my Mac to my phone. This let me add a few bells and whistles that a generic SSH client wouldn't have, like notifications when Claude Code was done working etc.
How did this work? The ban, I mean. Did you just wake up to find out an email and that your creds no longer worked? Were you doing things to sub-process out to the Claude Code CLI or something else?
I left a sibling comment detailing the technical side of things. I used the `Bun.spawn` API with the `terminal` key to give CC a PTY and mirrored it to my phone with xterm.js. I used SSE to stream CC data to xterm.js and a regular request to send commands out from my phone. In my mind, this is no different than using CC via SSH from my phone - I was still bound by the same limits and wasn't trying to bypass them, Anthropic is entitled to their different opinion of course.
And yeah, I got three (for some reason) emails titled "Your account has been suspended" whose content said "An internal investigation of suspicious signals associated with your account indicates a violation of our Usage Policy. As a result, we have revoked your access to Claude.". There is a link to a Google Form which I filled out, but I don't expect to hear back.
I did nothing even remotely suspicious with my Anthropic subscription so I am reasonably sure this mirroring is what got me banned.
Edit: BTW I have since iterated on doing the same mirroring using OpenCode with Codex, then Codex with Codex and now Pi with GPT-5.2 (non-Codex) and OpenAI hasn't banned me yet and I don't think they will as they decided to explicitly support using your subscription with third party coding agents following Anthropic's crackdown on OpenCode.
There is weaponized malaise employed by these frontier model providers and I feel like that dark-pattern, what you pointed out, and others are employed to rate-limit certain subscriptions.
They have two products:
* Subscription plans, which are (probably) subsidized and definitely oversubscribed (ie, 100% of subscribers could not use 100% of their tokens 100% of the time).
* Wholesale tokens, which are (probably) profitable.
If you try to use one product as the other product, it breaks their assumptions and business model.
I don't really see how this is weaponized malaise; capacity planning and some form of over-subscription is a widely accepted thing in every industry and product in the universe?
That's non-local. I am not interested in coding assistants that work on cloud based work-spaces. That's what motivated me to developed this feature for myself.
Yes, although OpenCode works great with official Claude API keys that are on normal API pricing.
What Anthropic blocked is using OpenCode with the Claude "individual plans" (like the $20/month Pro or $100/month Max plan), which Anthropic intends to be used only with the Claude Code client.
OpenCode had implemented some basic client spoofing so that this was working, but Anthropic updated to a more sophisticated client fingerprinting scheme which blocked OpenCode from using this individual plans.
Thank you for this comment! I knew it was something like this. I've been using it in the VSCode terminal, but you're right, the ANSI terminal just doesn't work. I wasn't quite sure why!
Did they actually say that? I thought they rolled it back.
OpenCode et al continue to work with my Max subscription.
OpenAI committed to allowing it btw. I don't know why Anthropic gets so much love here
Cause they make the best coding model.
It's that simple. Everyone else is trying to compete in other ways and Anthropic are pushing for dominate the market.
They'll eventually lose their performance edge and suddenly they will back to being cute and fluffy
I've cancelled a clause sub, but still have one.
Agreed.
I've tried all of the models available right now, and Claude Opus is by far the most capable.
I had an assertion failure triggered in a fairly complex open-source C library I was using, and Claude Opus not only found the cause, but wrote a self-contained reproduction code I could add to a GitHub issue. And it also added tests for that issue, and fixed the underlying issue.
I am sincerely impressed by the capabilities of Claude Opus. Too bad its usage is so expensive.
Because OpenAI is on the back foot at the moment, they need the retention
What do you require local models to do? The State of Utopia[1] is currently busy porting a small model to run in a zero-trust environment - your web browser. It's finished the port in javascript and is going to wasm now for the CPU path. you can see it being livecoded by Claude right now[2] (this is day 2, day 1 it ported the C++ code to javascript successfully). We are curious to know what permissions you would like to grant such a model and how you would like it served to you. (For example, we consider that you wouldn't trust a Go build - especially if it's built by a nation state, regardless of our branding, practices, members or contributors.)
Please list what capabilities you would like our local model to have and how you would like to have it served to you.
[1] a sovereign digital nation built on a national framework rather than a for-profit or even non-profit framework, will be available at https://stateofutopia.com (you can see some of my recent posts or comments here on HN.)
[2] https://www.youtube.com/live/0psQ2l4-USo?si=RVt2PhGy_A4nYFPi
> I’m not paying to support anticompetitive behaviour
You are doing that all the time. You just draw the line, arbitrarily.
What is the point of any of this? To exchange how we think about things. I think virtue signaling is boring and uncandid.
That's great, yes. We all draw the line somewhere, subjectively. We all pretend we follow logic and reason and lets all be more honest and truthfully share how we as humans are emotionally driven not logically driven.
It's like this old adage "Our brains are poor masters and great slaves". We are basically just wanting to survive and we've trained ourselves to follow the orders of our old corporate slave masters who are now failing us, and we are unfortunately out of fear paying and supporting anticompetitive behavior and our internal dissonance is stopping us from changing it (along with fear of survival and missing out and so forth).
The global marketing by the slave master class isn't helping. We can draw a line however arbitrary we'd like though and its still better and more helpful than complaining "you drew a line arbitrarily" and not actually doing any of the hard courageous work of drawing lines of any kind in the first place.
I got openclaw to compete Qwen3-Coder-Next vs Minimax M2.1 simultaneously on my Mac Studio 512GB: https://clutch-assistant.github.io/model-comparison-report/
I just tried qwen 3 tts and it was mind blowingly good, you can even provide directions for the overall tone etc. Which wasn't the case when I used commercial super expensive products like the (now closed after being bought by meta) play.ht .
Does anyone see a reason to still use elevenlabs etc. ?
> But my experience is they’re not really even close to the closed paid models.
They are usually as good as the flagship model for 12-18 months ago. Which may sound like a massive difference, because somehow it is, but it's also fairly reasonable, you don't need to live to the bleeding edge.
And it's worth pointing out that Claude Code now dispatches "subagents" from Opus->Sonnet and Opus->Haiku ... all the time, depending on the problem.
Running this thing locally on my Spark with 4-bit quant I'm getting 30-35 tokens/sec in opencode but it doesn't feel any "stupider" than Haiku, that's for sure. Haiku can be dumb as a post. This thing is smarter than that.
It feels somewhere around Sonnet 4 level, and I am finding it genuinely useful at 4-bit even. Though I have paid subscriptions elsewhere, so I doubt I'll actually use it much.
I could see configuration OpenCode somehow to use paid Kimi 2.5 or Gemini for the planning/analysis & compaction, and this for the task execution. It seems entirely competent.
Pretty cool that they are advertising OpenClaw compatibility. I've tried a few locally-hosted models with OpenClaw and did not get good results – (that tool is a context-monster... the models would get completely overwhelmed them with erroneous / old instructions.)
Granted these 80B models are probably optimized for H100/H200 which I do not have. Here's to hoping that OpenClaw compat. survives quantization
Not crazy about it. It keeps getting stuck in a loop and filling up the context window (131k, run locally). Kimi's been nice, even if a bit slow.
For someone who is very out of the loop with these AI models, can someone explain what I can actually run on my 3080ti (12G)? Is this something like that or is this still too big; is there anything remotely useful runnable with my GPU? I have 64G RAM if that helps (?).
This model does not fit in 12G of VRAM - even the smallest quant is unlikely to fit. However, portions can be offloaded to regular RAM / CPU with a performance hit.
I would recommend trying llama.cpp's llama-server with models of increasing size until you hit the best quality / speed tradeoff with your hardware that you're willing to accept.
The Unsloth guides are a great place to start: https://unsloth.ai/docs/models/qwen3-coder-next#llama.cpp-tu...
Thanks for the pointers!
one more thing, that guide says:
> You can choose UD-Q4_K_XL or other quantized versions.
I see eight different 4-bit quants (I assume that is the size I want?).. how to pick which one to use?
IQ4_XS
Q4_K_S
Q4_1
IQ4_NL
MXFP4_MOE
Q4_0
Q4_K_M
Q4_K_XLThe I-prefix stands for Imatrix smoothing in the quantization. It trades a little more accuracy for speed than other quant styles. The _0 and _1 quants are older, simpler quants that are very accurate but kinda slow. The K quants, in my limited understanding, primarily quantize at the specified bit depth, but will bump certain important areas higher, and less used parts lower. It generally performs better while providing similar accuracy to the _1 quants. MXFP4 is specific to Nvidia, so I can't use it on my AMD hardware. It's supposed to be very efficient. The UD part includes more of Unsloth's speed optimizations.
Also, depending on how much regular system RAM you have, you can offload mixture-of-expert models like this, keeping only the most important layers on your GPU. This may let you use larger, more accurate quants. That is functionality that is supported by llama.cpp and other frameworks and is worth looking into how to do.
This model is exactly what you’d want for your resources. GPU for prompt processing, ram for model weights and context length, and it being MoE makes it fairly zippy. Q4 is decent; Q5-6 is even better, assuming you can spare the resources. Going past q6 goes into heavily diminishing resources.
This GGUF is 48.4GB - https://huggingface.co/Qwen/Qwen3-Coder-Next-GGUF/tree/main/... - which should be usable on higher end laptops.
I still haven't experienced a local model that fits on my 64GB MacBook Pro and can run a coding agent like Codex CLI or Claude code well enough to be useful.
Maybe this will be the one? This Unsloth guide from a sibling comment suggests it might be: https://unsloth.ai/docs/models/qwen3-coder-next