Also of interest: DeepSeek V4-Pro-0813 (1.6T-A49B) benchmark scores have apparently just been announced on the DeepSeek WeChat channel and they're sitting about Fable 5 level.[1]
This may not 'quite a bit behind' those at all. If you look at the benchmark numbers they are very comparable to Fable, but beyond a certain point the benchmark numbers don't tell you much. Opus #5 beats Fable on some benchmarks but given similar cost almost everyone who has used those two models will prefer to use Fable.
At this price range $0.87per 1M they will get a lot of usage of people trying it out. Given the benchmark numbers, for many people and many use cases this will become their primary driver. There are people and use cases where Fable, Sol will work better but those are likely not the target of DeepSeek anyway.
In terms of performance and price pareto curve I don't think any model can beat this today (though openAI is doing some exciting recent work in efficiency) - which is a remarkable feat for the DeepSeek team.
Either way, what a time for consumers of these models :)
Supposedly this is a Kimi k3 rival. Bit of a chonker, especially since they only released bf16 and fp8. So at launch this will be harder to serve than k3. No QAT on q4 means that someone with deep pockets (nvda?) will have to quant it, with plenty of calibration data. Should bring it ~1.3TB, so around k3 size.
License pretty similar to k3 with some caveats. Free to use for internal or <50M$ revenue / year. Limitations above that threshold for serving the model or services targeting coding / productivity agents.
Benchmarks are looking good, trading blows w/ opus4.8 and sol, generally 10-20p under fable. But that's neither here nor there w/ qwen, their benchmark to real world usage correlation has been iffy in the past.
The local model 3.8-27B announced for Friday, same time so ~48 hours from now. That'll be a bit more exciting for a lot more people, since 3.6 was quite good for local inference, and their 3.7-max -> 3.8-max shows a lot of improvement.
3.8-27B is confirmed for Friday. They didn't release their whole 4B-400B range of models since 3.5. And 3.6 only got 27B and 35B MoE. So yeah, slowing down, but not completely out of the small model game.
unshloth started as a finetuning library with lots of optimisations so you could finetune on lower end hardware. Kind of OGs of the local community. Started by two brothers Michael and Daniel(?) a math wiz and a community builder/communicator. They've since gotten some VC backing, are active in quantising lots of models on release day (work w/ labs to prepare things), known for their optimised quants (use different bits for different layers). Recently I saw they launched some sort of a desktop app, like lmstudio if you're familiar with it. They're really cool people and known in the local model places.
They started with offering training methods for quantized models to save memory and added new things over time. They are very active in the local model community and have extensive documentation and tooling to help with running and training models locally.
Llama.cpp can quantize without special training, but I'm not sure if any special model architecture support is needed to read it in the first place. If it can be converted to gguf at all and you know what tensors to target, it can get the full ternary bonsai treatment today.
Sure, but that's for "personal" serving. I meant for 3rd party providers. Usually we get a good indication on what it costs to host this, as the prices settle on open router. That's why I said it's tougher to serve than kimi k3 on launch. As a provider you'd do fp8 if the model creator didn't do QAT on q4, or until someone does a good calibrated nvfp4. And that's usually nvda :)
quanting is actually cheap and you can compress a model that does not fit on a GPU. You can process layer by layer, this is what the sequential processor in llm-compressor does.
The 1bit quant model is at an astonishing 397GB with 95B active per MOE. This literally puts Opus 4.5 performance level into a machine a normal person could buy, and still gets usable tokens/second.
The full lossless model BF16 is clocking at 4.9TB. The model card claims the model to be between Opus 4.8 and Fable 5. Again that's astonishing as getting a machine with 7TB RAM (with context + KV cache) is still within the realm of medium size companies.
Bad things: The open source version has its vision capability removed, and the context capped at 250k . I expect someone to bolt a Kimi 2.6 vision tower to it to restore the vision capability (at less performance of course). For context, I played around with extending the context to 600k for Qwen 3.5 397b, and the context remained stable up to around 480k. It'd be interesting to see if the same can be done to Q3.8 .
Also no out of the box DSpark/DFlash support. MTP is present so we should at least get some boost in TP speed.
Extremely large 1 bit models are usually within 50-60% of KV divergence to lossless models. In this case I think the comparison to Opus 4.5 is a fair assessment.
Extremely large models don't suffer as much from quantization due to its weight topology also contains encoded information, so the loss of info from any one weight is somewhat mitigated.
Any one weight, but all of them. And also crushing the architecture itself?
I wouldn't pick up 400gb of hardware to run in that mode. I might try it for fun, but even then you are looking at handling a 95GB active parameter set.
This is NOT a model for most home labs. I'm sure some can and will use it. But most, should steer clear.
KL divergence (you misspelled it) doesn't tell you anything about capability drop - how much did this particular benchmark (thus ranking among models) change after 10% or 50% KL divergence?
Opus 4.5 level of performance is also accessible with deepseek-v4-flash-0731 (0731 being the july 31 update) which is much, much, much smaller. 2x RTX pro 6000 blackwell can run it. 4x can run it very comfortably
I am running DS v4 flash 0731 lossless at 80t/s right now. It really is not at Opus 4.5 level (for my workload). I would say it's around 3.7 Sonnet, which is still pretty good, but other models such as GLM 5.2 are still leaps better. Of course I run DSv4 flash over GLM 5.2 for a few very good reasons, but intelligence is not 1 of them.
Despite fitting into VRAM, I can't get DSV4 to run at usable speeds on my AMD hardware. The upcoming qwen3.8 27b greatly excites me, and I hope it can outperform Stepfun 3.7 Flash, which is the best thing I can run today.
I was under the impression that you could fit the full 1M context within the 192GB VRAM as a result of DeepSeek's various architectural advancements, but I'll grant that DSpark + a larger pool for concurrency may necessitate more VRAM, yes.
I don't understand the logic behind model sizes and quantization.
Suppose I have 100GB of unified memory, how should I know which model suits it best? I understand how a 2.4T model wouldn't fit, but I don't understand the impact of quantization and whether I should use a 200G model quantised to fit say 90GB of memory, or a non-quantised 90G model.
It really depends. It used to be easier to have a rule of thumb, but now it's not clear anymore. Now there are a lot of things to consider, such as a model's kv efficiency (how much context you can fit), MoE v. dense, QAT or not (Quant aware training) and so on.
The old rule of thumb was that a lower quant of a larger model > higher quant of a smaller model. That being said, for some things going lower than fp8 will see a lot of degradation in generation quality. Except if the model comes with QAT 4bit quants. Then there's also nvfp4 w/ calibration data, which also can improve things. So it's really not easy to tell "at a glance" you'd have to test them yourself on your hardware.
Standard models are designed to quantize down to 4-bits relatively well.
Anything below that, and especially 1.58b - is typically complete garbage, and you're much better off running a model 100x smaller at regular precision (compared to one 7x smaller quantized into complete garbage).
If the model was designed specifically to quantize down to 1.58b, then it's different.
AFAIK, there's no large models designed for this yet.
> In particular, Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools, etc.
That is unfortunate, that the open weight model doesn't have vision support or the 1M context length...
People have had surprising success adding vision to open-weight LLMs that ship without it, like DSV4 Flash [1] or GLM-5.2 [2]. Given this model is already vision-trained I expect that approach will work well here.
I've been wanting to run open weight models lately to give them a shot with OpenCode. However, I get the impression that models like Qwen and Kimi k3 are impossible to run locally? I have a RTX 5090 and 64 GB of RAM but the models seem to be much larger than that. What's the route to start using these models? Bedrock?
OpenRouter is (roughly) a single proxy between you + many different models + providers. it works with opencode (+ many other products), and is relatively convenient for trying out a bunch of models.
note that they add some fee ontop of things (maybe 10% of spend?). it isn't htat big of a deal for general experimentation, but if you end up wanting to use a single model in a higher-volume way, it likely makes sense to cut them out of your stack.
Bedrock seems to have stopped adding new open-weights models, and mostly only has Anthropic and OpenAI stuff now. You can get Qwen 3.8 directly from Alibaba: https://www.qwencloud.com (proprietary variant) or from DigitalOcean (this variant, probably also from others soon).
On your 5090 you could easily run a smaller model like Qwen 3.6 27B: https://huggingface.co/collections/Qwen/qwen36 or Gemma 4 etc., or as mentioned there's a Qwen 3.8 27B coming out in a few days.
No, Kimi k3 is 2.8T params. This is 2.4T params but ~5TB weights because it was released in bf16 and ~2.5TB for the fp8 version. Kimi k3 launched with QAT 4bit, so ~1.5TB weights.
Also of interest: DeepSeek V4-Pro-0813 (1.6T-A49B) benchmark scores have apparently just been announced on the DeepSeek WeChat channel and they're sitting about Fable 5 level.[1]
[1] https://www.reddit.com/r/LocalLLaMA/comments/1vmi0fg/deepsee...
Isn't this quite a bit behind Sol and Fable and even ChatGPT 5.5 xhigh and Opus 5 max?
In terms of what you get for what you pay for, it's incredible - probably by far the best.
But unless I'm reading things wrong, it does not appear to be top-of-the-line.
This may not 'quite a bit behind' those at all. If you look at the benchmark numbers they are very comparable to Fable, but beyond a certain point the benchmark numbers don't tell you much. Opus #5 beats Fable on some benchmarks but given similar cost almost everyone who has used those two models will prefer to use Fable.
At this price range $0.87per 1M they will get a lot of usage of people trying it out. Given the benchmark numbers, for many people and many use cases this will become their primary driver. There are people and use cases where Fable, Sol will work better but those are likely not the target of DeepSeek anyway.
In terms of performance and price pareto curve I don't think any model can beat this today (though openAI is doing some exciting recent work in efficiency) - which is a remarkable feat for the DeepSeek team.
Either way, what a time for consumers of these models :)
And relatedly, just now available on OpenRouter
https://openrouter.ai/deepseek/deepseek-v4-pro-0813
Supposedly this is a Kimi k3 rival. Bit of a chonker, especially since they only released bf16 and fp8. So at launch this will be harder to serve than k3. No QAT on q4 means that someone with deep pockets (nvda?) will have to quant it, with plenty of calibration data. Should bring it ~1.3TB, so around k3 size.
License pretty similar to k3 with some caveats. Free to use for internal or <50M$ revenue / year. Limitations above that threshold for serving the model or services targeting coding / productivity agents.
Benchmarks are looking good, trading blows w/ opus4.8 and sol, generally 10-20p under fable. But that's neither here nor there w/ qwen, their benchmark to real world usage correlation has been iffy in the past.
The local model 3.8-27B announced for Friday, same time so ~48 hours from now. That'll be a bit more exciting for a lot more people, since 3.6 was quite good for local inference, and their 3.7-max -> 3.8-max shows a lot of improvement.
> 3.8-27B announced for Friday
Maybe I’m misreading this or some other post, I thought QWEN was stepping away from releasing these models for local consumption
3.8-27B is confirmed for Friday. They didn't release their whole 4B-400B range of models since 3.5. And 3.6 only got 27B and 35B MoE. So yeah, slowing down, but not completely out of the small model game.
They reversed course and now are saying they'll be releasing their Max style models in open weights.
thank you china!
Unsloth already has a guide for their quants: https://unsloth.ai/docs/models/qwen3.8
I wonder who is unsloth and where they got time, hardware and knowledge to quantize them?
Daniel Han is just that good!
unshloth started as a finetuning library with lots of optimisations so you could finetune on lower end hardware. Kind of OGs of the local community. Started by two brothers Michael and Daniel(?) a math wiz and a community builder/communicator. They've since gotten some VC backing, are active in quantising lots of models on release day (work w/ labs to prepare things), known for their optimised quants (use different bits for different layers). Recently I saw they launched some sort of a desktop app, like lmstudio if you're familiar with it. They're really cool people and known in the local model places.
It's our guy Daniel: https://www.linkedin.com/in/danielhanchen
They started with offering training methods for quantized models to save memory and added new things over time. They are very active in the local model community and have extensive documentation and tooling to help with running and training models locally.
Llama.cpp can quantize without special training, but I'm not sure if any special model architecture support is needed to read it in the first place. If it can be converted to gguf at all and you know what tensors to target, it can get the full ternary bonsai treatment today.
Sure, but that's for "personal" serving. I meant for 3rd party providers. Usually we get a good indication on what it costs to host this, as the prices settle on open router. That's why I said it's tougher to serve than kimi k3 on launch. As a provider you'd do fp8 if the model creator didn't do QAT on q4, or until someone does a good calibrated nvfp4. And that's usually nvda :)
That makes sense, but your specific phrasing precluded the possibility of non-QAT quantization.
Should have worded that better, my bad.
QAT is an optimizing quantization algorithm, not naive quant.
Right, but the way they phrased it suggested that without QAT it could not be quanted at all.
Now that they have reached the frontier in raw performance, I would like to see Chinese models improve their reasoning efficiency.
For all the talk about over reasoning, K3 on low thinking has been rather nice
quanting is actually cheap and you can compress a model that does not fit on a GPU. You can process layer by layer, this is what the sequential processor in llm-compressor does.
https://unsloth.ai/docs/models/qwen3.8
The 1bit quant model is at an astonishing 397GB with 95B active per MOE. This literally puts Opus 4.5 performance level into a machine a normal person could buy, and still gets usable tokens/second.
The full lossless model BF16 is clocking at 4.9TB. The model card claims the model to be between Opus 4.8 and Fable 5. Again that's astonishing as getting a machine with 7TB RAM (with context + KV cache) is still within the realm of medium size companies.
Bad things: The open source version has its vision capability removed, and the context capped at 250k . I expect someone to bolt a Kimi 2.6 vision tower to it to restore the vision capability (at less performance of course). For context, I played around with extending the context to 600k for Qwen 3.5 397b, and the context remained stable up to around 480k. It'd be interesting to see if the same can be done to Q3.8 .
Also no out of the box DSpark/DFlash support. MTP is present so we should at least get some boost in TP speed.
To compare a 1 bit quant to the full fat model is misleading.
Honestly this model people at home can tinker with, if you have a big enough Mac. Maybe 4 Strix Halo/DGX Spark, and then at 1 bit quant? Nah.
Use the right sized model, for your hardware. You'll get better results.
Extremely large 1 bit models are usually within 50-60% of KV divergence to lossless models. In this case I think the comparison to Opus 4.5 is a fair assessment.
Extremely large models don't suffer as much from quantization due to its weight topology also contains encoded information, so the loss of info from any one weight is somewhat mitigated.
Any one weight, but all of them. And also crushing the architecture itself?
I wouldn't pick up 400gb of hardware to run in that mode. I might try it for fun, but even then you are looking at handling a 95GB active parameter set.
This is NOT a model for most home labs. I'm sure some can and will use it. But most, should steer clear.
KL divergence (you misspelled it) doesn't tell you anything about capability drop - how much did this particular benchmark (thus ranking among models) change after 10% or 50% KL divergence?
Opus 4.5 level of performance is also accessible with deepseek-v4-flash-0731 (0731 being the july 31 update) which is much, much, much smaller. 2x RTX pro 6000 blackwell can run it. 4x can run it very comfortably
I am running DS v4 flash 0731 lossless at 80t/s right now. It really is not at Opus 4.5 level (for my workload). I would say it's around 3.7 Sonnet, which is still pretty good, but other models such as GLM 5.2 are still leaps better. Of course I run DSv4 flash over GLM 5.2 for a few very good reasons, but intelligence is not 1 of them.
Despite fitting into VRAM, I can't get DSV4 to run at usable speeds on my AMD hardware. The upcoming qwen3.8 27b greatly excites me, and I hope it can outperform Stepfun 3.7 Flash, which is the best thing I can run today.
What do you need the extra 2 for? Tensor parallelism?
Longer context and more cache. The problem is that native format with DSpark enabled you have very little room on the VRAM.
I was under the impression that you could fit the full 1M context within the 192GB VRAM as a result of DeepSeek's various architectural advancements, but I'll grant that DSpark + a larger pool for concurrency may necessitate more VRAM, yes.
I don't understand the logic behind model sizes and quantization.
Suppose I have 100GB of unified memory, how should I know which model suits it best? I understand how a 2.4T model wouldn't fit, but I don't understand the impact of quantization and whether I should use a 200G model quantised to fit say 90GB of memory, or a non-quantised 90G model.
It really depends. It used to be easier to have a rule of thumb, but now it's not clear anymore. Now there are a lot of things to consider, such as a model's kv efficiency (how much context you can fit), MoE v. dense, QAT or not (Quant aware training) and so on.
The old rule of thumb was that a lower quant of a larger model > higher quant of a smaller model. That being said, for some things going lower than fp8 will see a lot of degradation in generation quality. Except if the model comes with QAT 4bit quants. Then there's also nvfp4 w/ calibration data, which also can improve things. So it's really not easy to tell "at a glance" you'd have to test them yourself on your hardware.
Standard models are designed to quantize down to 4-bits relatively well.
Anything below that, and especially 1.58b - is typically complete garbage, and you're much better off running a model 100x smaller at regular precision (compared to one 7x smaller quantized into complete garbage).
If the model was designed specifically to quantize down to 1.58b, then it's different.
AFAIK, there's no large models designed for this yet.
Usually 4-bit 200B model is better than 8-bit 90B. But if you go below 4 bits, I am not sure what is better.
There's no rhyme or reason to it. Quants aren't benchmarked much. Generally 4bit better than smaller model 8bit
> The 1bit quant model i
at this kind of quantization is it useful though?
> In particular, Qwen3.8-Max is the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools, etc.
That is unfortunate, that the open weight model doesn't have vision support or the 1M context length...
People have had surprising success adding vision to open-weight LLMs that ship without it, like DSV4 Flash [1] or GLM-5.2 [2]. Given this model is already vision-trained I expect that approach will work well here.
[1] https://old.reddit.com/r/LocalLLaMA/comments/1vl6ior/i_gave_...
[2] https://huggingface.co/baseten/GLM-5.2-Vision-NVFP4
Qwen3.5 was awesome: fairly open and fully featured. 3.8 lacking vision, nerfing thinking modes, and low context length feels pointless.
I've been wanting to run open weight models lately to give them a shot with OpenCode. However, I get the impression that models like Qwen and Kimi k3 are impossible to run locally? I have a RTX 5090 and 64 GB of RAM but the models seem to be much larger than that. What's the route to start using these models? Bedrock?
5090 is plenty for the Q4_K_M quantized version of 3.6 27B with reduced context size.
I run it on a 3090(24GB) and 64k context using GGUF format and llama-cpp. Double 3090 gives you 128k, quad 3090 gets you to full context - 256k.
OpenRouter is (roughly) a single proxy between you + many different models + providers. it works with opencode (+ many other products), and is relatively convenient for trying out a bunch of models.
for example, they already have qwen3.8-max
https://openrouter.ai/discover?model=qwen/qwen3.8-max
note that they add some fee ontop of things (maybe 10% of spend?). it isn't htat big of a deal for general experimentation, but if you end up wanting to use a single model in a higher-volume way, it likely makes sense to cut them out of your stack.
Bedrock seems to have stopped adding new open-weights models, and mostly only has Anthropic and OpenAI stuff now. You can get Qwen 3.8 directly from Alibaba: https://www.qwencloud.com (proprietary variant) or from DigitalOcean (this variant, probably also from others soon).
On your 5090 you could easily run a smaller model like Qwen 3.6 27B: https://huggingface.co/collections/Qwen/qwen36 or Gemma 4 etc., or as mentioned there's a Qwen 3.8 27B coming out in a few days.
What does the number before the B signify?
Its number of parameters. The 'B' is billions. If you have bf16 weights each parameter would be 16 bits.
You could easily run any of their 30B-or-less models which is what most people are waiting for.
Apparently the ~30B variant will be released on Friday?
Fireworks or OpenCode Go
More curious about how qwen3.8-27B performs. That's the size that I can run locally.
Is this the largest ever open weight model release by parameter count? I think it is.
No, Kimi k3 is 2.8T params. This is 2.4T params but ~5TB weights because it was released in bf16 and ~2.5TB for the fp8 version. Kimi k3 launched with QAT 4bit, so ~1.5TB weights.
KIMI K3 was the biggest open weight release afaik; It is 2.8T-A100B if I'm correct
Do we know if AA and DeepSWE benchmarks are on bf16 or fp8 quantisations?
"QwenSVGBench" elo 1713, pelicanmaxxxing confirmed?
The card looks almost too good to be true
Not seeing the upside versus K3 here, especially with the intentional capability loss.
Read the room, Qwen. It's not a good time to hobble your releases.
A ~5TB model.
the a little disappointing part is this is released in BF16. so i suppose no QAT was implemented.