I started with antirez' DwarfStar[1] on one spark and that (~11-14tok/s generation, ~300-400 tok/s prompt processing) was enough of a taste for me to jump into 2 sparks, running the native quant of DSv4 Flash.
Now at 40-50tok/s generation and ~2000 tok/s prefill with a model that I've seen reason through race conditions and be able to trivially pull off any straight-forward coding task, and remain coherent at 500k context. With a preview checkpoint of the weights!
I'm excited for the future of local LLMs. There is some buy-in but apparently not an extreme amount to get access to models that can stand in the for the giants on all but the most challenging and/or hands-off coding tasks.
I suspect DwarfStar could probably squeeze more performance out of the single spark, maybe up closer to 20tok/s.
Moving to 2 sparks meant switching to vLLM with 2-way tensor parallelism and working multi-token prediction. The parallelism and MTP on top of better tuned kernels[1] gave an extremely nice boost! I was quite pleased. I've seen bursts up to 60tok/s at ~150k context - sometimes the MTP seems to really kick in (i.e. high acceptance rate on its tokens)
Currently running a custom vLLM build put together by some folks on the Nvidia forums[2], which speaks to how early support for the model is.
Personally, I've tried to squeeze more tok/s for a single DGX Spark deployment and DeepSeek V4 Flash but only got marginal improvements. There's work to do on fusing kernels and other optimizations that are already on antirez's roadmap so it is not worth duplicating efforts.
I've had positive experiences running GLM 4.7 via vLLM, tool calling works well and the inference is fast. Do you run DeepSeek V4 Flash on vLLM?
DeepSeek v4 Flash MTP is a training optimization. It doesn't make inference run faster, it must run the entire model forward as the "verifier." This is in the paper, and this is why the docs they release do not mention using it for accelerated inference.
Eventually, I'm going to stop writing stuff like this @dang, because even though it is literally being read by a human, it's going to just be copy and pasted into a chatbot, which will actually spend the time trying to comprehend what I am saying.
> MTP in Inference. Our MTP strategy mainly aims to improve the performance of the main model, so during inference, we can directly discard the MTP modules and the main model can function independently and normally. *Additionally, we can also repurpose these MTP modules for speculative decoding to further improve the generation latency.*[1]
(emphasis mine)
> Instead of predicting just the next single token, DeepSeek-V3 predicts the next 2 tokens through the MTP technique. Combined with the framework of speculative decoding (Leviathan et al., 2023; Xia et al., 2023), it can significantly accelerate the decoding speed of the model.[2]
> As DeepSeek-V3, DeepSeek-V4 series also set MTP modules and
objectives. Given that the MTP strategy has been validated in DeepSeek-V3, we adopt the same strategy for DeepSeek-V4 series without modification.[3]
I’ve been considering a move to local llm setup, having been underwhelmed coat vs value of various online offerings. But at the same time worried anything I get will be obsolete in a couple months. And I don’t want to have to babysit it. I really want some agents managing and creating side hustles for me and have some other things. I’m technical-have written my own harness and use gh copilot and grok daily and have a hosted openwebui+openrouter thing. I’m also torn between a 128g MacBook Pro or a framework, or spark or similar and lightweight laptop to access. Would love advice anyone has for (or against) going local. I have asked ai but have analysis paralysis as 5k would be a big investment for me so I want to make right choices
Well, if you are making side-hustle money now using online models that, critically, you could also run at home, then it sounds like it’s just a matter of numbers. Oh and, unless you spend a lot more than 5k, your local model will still be slower than the online model. What’s your estimated ROI?
Assuming that’s not true based on your phrasing, you’d be shooting yourself in the foot. Start using online models with the same quant at least benchmark as what you could run at home. Prepare for the at home model to be slower.
Mac, DGX Spark, and a Framework Desktop / Ryzen AI Max 395 (ie Strix Halo) will not give you great performance running LLMs. One benefit of the Spark over the others is you can easily link up to 4 of them. Only MoE (sparse) models will be usable. Even if you can run some massive models, they will crawl. You're better off running one or more GPU cards.
You probably want to try renting some time on a dedicated box with roughly the specs you’re considering and running the open models for a bit to see if you would actually use them before dropping a lot on local hardware. A 128 gig MacBook Pro isn’t going to get you an amazing model, and certainly not amazing speed. GLM 5.2 wants something like 350+ gigs at fp4 iirc.
I ran glm 5.2 on rented 8x h200 it could only do 2x concurrency at a cost of $40 an hour. It felt great but dang I wish it was cheaper... It needs 750 at fp8
> You probably want to try renting some time on a dedicated box with roughly the specs you’re considering and running the open models
You don't even need to go that far. For example, with Exoscale Dedicated Inference[1] you just point it at the Hugging Face for the model and quantisation you want to test and it automagically spits out an OpenAI-compatible API endpoint.
> I think they're just suggesting renting as a way to test
Well, yes, I understood that.
Which is why I started with the words "You don't even need to go that far.".
To re-phrase what I said in clearer terms:
Instead of renting an instance, then messing around with configuring Linux and whatever via SSH or Ansible or whatever. Just point a Hugging Face link at this magic service and get a ready-to-go API back. Enabling you to test your desired model spec with minimum fuss.
Ultimately the guy wants his own hardware. So why waste time messing around with someone else's VM if you just want to test a specific model spec. That is the TL;DR.
Have you tried llama.cpp with unsloth and models suited to it? GLM flash? It seemed to allow more models to be tried soon after they are released. Haven’t tried for long term deployment though, that’s the next step.
Highy anecdotal: I have tried various self-hosted models using both vllm and llama.cpp. I am in a situation where I have access to large amount of memory (~320 GB).
While experimenting with quantization I found that there is a non-trivial tradeoff between quality and memory footprint. Overall my experience follows the reported pattern of "2-bit is mwah, 4-bit half decent and 6-bit required for programming. Still, although MiniMax-m2.7 is useable with the 6-bit quantizations that unsloth provides, it felt like such a breath of fresh air when I used the reference full-size model.
I find it difficult to say why. I had mostly the same setup as before (parsing had to be slightly adjusted in Zed). Aside from not experiencing the thinking loops (where minimax would get stuck generating the same sentences over and over) there is little evidence of any real improvement (although the average thinking time felt shorter).
I would recommend against very low quantizations of GLM 5.0/5.1/5.2 or Kimi 2.5/2.6. Smaller models were more reliable, and therefore more useful.
I only have access to 96GB VRAM locally, but I'd agree with the general approach of avoiding lower quantizations, often anything below Q8 seems to suffer greatly on quality and seemingly never worth going below it, better to go for smaller model in that case.
With the exception of DwarfStar + DS4-Flash with IQ2_XXS quantization, which somehow seems to not suffer as much as I'd thought. I'd still opt for a smaller model + at least Q8.
I have tried llama-cpp, vllm is nicer (ray, handles queueing, doesn't have the cache invalidation bug for qwen/gemma models) and unsloth has toxic employees in their discord.
I've run 2 qwen/gemma @8bit with full context window side-by-side. Right now I have 4 models on my spark (qwen36moe, embedding, reranker, qwen3-1.7B) to support my markdown kb tool.
The setup is not as capable, but still good and gets better with models/algos. To me, it's more about the freedom to tinker, freedom from token bill anxiety, and potential right to compute should the government/oligarchy decides it gets to decide who can access which models.
I shared a project in their #research channel where I used their qwen36moe quant to refresh my PhD research. The channel had a topic that ended with something like "and all things research..."
One of their people accused me of self-promotion, and I reiterated that I shared it in that channel because it was their quant doing something (I thought) interesting as a research model. The number of people interested in the topic can be counted on your hands (in binary).
They remained accusatory, made it personal, and then started deleting messages. I suppose I escalated a bit (from their perspective), saying how this was not a good first encounter, they could have asked me to move it instead of just deleting it. Then they deleted every message, including all of their own, and put me in timeout. Erased from history, unable to participate, and so I left.
A coworker of mine (ML guy) is also sus about their quants, not nefarious, more that their benchmark results do not mean they are better, possibly skewed / benchmaxxed.
My impression is that with the latest round of high-profile releases, the open-weight "market" is coalescing around two players, DS4 Flash for speed and GLM 5.2 for smarts. Qwen is being left behind to pick up the scraps for the terminally GPU-poor.
We know they have what it takes to fight back, and they know it... so I agree, there's no reason not be optimistic about future Qwen releases. But then I've never really understood what motivates these releases in the first place.
Author here. Quick context the post doesn't quite spell out:
The tool_choice="auto" failure on Qwen3-Next isn't a parser issue — the model
reasons inside <think>, decides, and never emits the tool call. No error, just
empty tool_calls. The fix was swapping the backbone from Thinking to Instruct,
not tuning any parser flag.
The "load the bigger model first, size the smaller against actual residency"
playbook generalizes to anything with shared CUDA framework overhead. The ~5 GiB
framework floor shows up even at small gpu_memory_utilization values — plan
against actuals, not targets.
(...)
- Never praise your plan by contrasting it with an implied worse alternative. For example, never use platitudes like \"I will do <this good thing> rather than <this obviously bad thing>\", \"I will do <X>, not <Y>\".
- Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query.
(...)
```
It seems the OpenAI people added that first bullet to specifically address the tendency the model has, as seen in the parent comment. The goblin stuff coincidentally appears right after in the system prompt, so in included it as a bonus.
FWIW the mere fact that the goblin stuff is necessary and is in the system prompt suggests to me that OpenAI's approach of training ridiculously large models that can do everything for everyone is hopelessly cooked.
Though I concede it is not that much different than straightening the tie of your most valuable employee before you unwisely put them in front of a client and saying "please don't tell them about the regressions they didn't notice and remember, they don't want things explained in allegories drawn from the Silmarillion".
Can you try and tune your Claude or whatever LLM you're using for your text to phrase things in plain English. Way less use of antithesis, at least. You can probably find a skill for it, if not get an LLM to write your own.
I started with antirez' DwarfStar[1] on one spark and that (~11-14tok/s generation, ~300-400 tok/s prompt processing) was enough of a taste for me to jump into 2 sparks, running the native quant of DSv4 Flash.
Now at 40-50tok/s generation and ~2000 tok/s prefill with a model that I've seen reason through race conditions and be able to trivially pull off any straight-forward coding task, and remain coherent at 500k context. With a preview checkpoint of the weights!
I'm excited for the future of local LLMs. There is some buy-in but apparently not an extreme amount to get access to models that can stand in the for the giants on all but the most challenging and/or hands-off coding tasks.
[1]: https://github.com/antirez/ds4
> Now at 40-50tok/s generation and ~2000 tok/s
Not clear how you went from ~11-14 to ~40-50 tok/s. Is it by running the quant native model and adding a second Spark?
Cheers
Just chiming in - the claims above are real, I have very similar numbers in a cluster of 2x GX10 I have access to.
Instructions to reproduce, and benchmarks here: https://forums.developer.nvidia.com/t/deepseek-v4-flash-offi...
I suspect DwarfStar could probably squeeze more performance out of the single spark, maybe up closer to 20tok/s.
Moving to 2 sparks meant switching to vLLM with 2-way tensor parallelism and working multi-token prediction. The parallelism and MTP on top of better tuned kernels[1] gave an extremely nice boost! I was quite pleased. I've seen bursts up to 60tok/s at ~150k context - sometimes the MTP seems to really kick in (i.e. high acceptance rate on its tokens)
Currently running a custom vLLM build put together by some folks on the Nvidia forums[2], which speaks to how early support for the model is.
[1]: https://github.com/lukealonso/b12x
[2]: https://forums.developer.nvidia.com/t/372268
Personally, I've tried to squeeze more tok/s for a single DGX Spark deployment and DeepSeek V4 Flash but only got marginal improvements. There's work to do on fusing kernels and other optimizations that are already on antirez's roadmap so it is not worth duplicating efforts.
I've had positive experiences running GLM 4.7 via vLLM, tool calling works well and the inference is fast. Do you run DeepSeek V4 Flash on vLLM?
Yep, those are the numbers I'm getting with DSv4 Flash on vLLM across 2 sparks.
DeepSeek v4 Flash MTP is a training optimization. It doesn't make inference run faster, it must run the entire model forward as the "verifier." This is in the paper, and this is why the docs they release do not mention using it for accelerated inference.
Eventually, I'm going to stop writing stuff like this @dang, because even though it is literally being read by a human, it's going to just be copy and pasted into a chatbot, which will actually spend the time trying to comprehend what I am saying.
> MTP in Inference. Our MTP strategy mainly aims to improve the performance of the main model, so during inference, we can directly discard the MTP modules and the main model can function independently and normally. *Additionally, we can also repurpose these MTP modules for speculative decoding to further improve the generation latency.*[1]
(emphasis mine)
> Instead of predicting just the next single token, DeepSeek-V3 predicts the next 2 tokens through the MTP technique. Combined with the framework of speculative decoding (Leviathan et al., 2023; Xia et al., 2023), it can significantly accelerate the decoding speed of the model.[2]
> As DeepSeek-V3, DeepSeek-V4 series also set MTP modules and objectives. Given that the MTP strategy has been validated in DeepSeek-V3, we adopt the same strategy for DeepSeek-V4 series without modification.[3]
[1]: https://arxiv.org/pdf/2412.19437#subsection.2.2
[2]: https://arxiv.org/pdf/2412.19437#subsubsection.5.4.3
[3]: https://arxiv.org/pdf/2606.19348v1#subsection.2.1
Side comment: I feel you may be too cynical towards your fellow commenters.
I’ve been considering a move to local llm setup, having been underwhelmed coat vs value of various online offerings. But at the same time worried anything I get will be obsolete in a couple months. And I don’t want to have to babysit it. I really want some agents managing and creating side hustles for me and have some other things. I’m technical-have written my own harness and use gh copilot and grok daily and have a hosted openwebui+openrouter thing. I’m also torn between a 128g MacBook Pro or a framework, or spark or similar and lightweight laptop to access. Would love advice anyone has for (or against) going local. I have asked ai but have analysis paralysis as 5k would be a big investment for me so I want to make right choices
Well, if you are making side-hustle money now using online models that, critically, you could also run at home, then it sounds like it’s just a matter of numbers. Oh and, unless you spend a lot more than 5k, your local model will still be slower than the online model. What’s your estimated ROI?
Assuming that’s not true based on your phrasing, you’d be shooting yourself in the foot. Start using online models with the same quant at least benchmark as what you could run at home. Prepare for the at home model to be slower.
no one is making money side-hustling ai models. This is like reddit wet dream. get real, dont get scammed by ppl selling you these dreams.
Mac, DGX Spark, and a Framework Desktop / Ryzen AI Max 395 (ie Strix Halo) will not give you great performance running LLMs. One benefit of the Spark over the others is you can easily link up to 4 of them. Only MoE (sparse) models will be usable. Even if you can run some massive models, they will crawl. You're better off running one or more GPU cards.
You probably want to try renting some time on a dedicated box with roughly the specs you’re considering and running the open models for a bit to see if you would actually use them before dropping a lot on local hardware. A 128 gig MacBook Pro isn’t going to get you an amazing model, and certainly not amazing speed. GLM 5.2 wants something like 350+ gigs at fp4 iirc.
I ran glm 5.2 on rented 8x h200 it could only do 2x concurrency at a cost of $40 an hour. It felt great but dang I wish it was cheaper... It needs 750 at fp8
what was the concurrency limitation? that node should be able to support a lot more
> You probably want to try renting some time on a dedicated box with roughly the specs you’re considering and running the open models
You don't even need to go that far. For example, with Exoscale Dedicated Inference[1] you just point it at the Hugging Face for the model and quantisation you want to test and it automagically spits out an OpenAI-compatible API endpoint.
[1] https://www.exoscale.com/ai-cloud-infrastructure/dedicated-i...
(I have no relationship with Exoscale, this particular product just crossed my radar recently)
I think they're just suggesting renting as a way to test that the hardware they're considering purchasing would actually be able to do what they need.
> I think they're just suggesting renting as a way to test
Well, yes, I understood that.
Which is why I started with the words "You don't even need to go that far.".
To re-phrase what I said in clearer terms:
Instead of renting an instance, then messing around with configuring Linux and whatever via SSH or Ansible or whatever. Just point a Hugging Face link at this magic service and get a ready-to-go API back. Enabling you to test your desired model spec with minimum fuss.
Ultimately the guy wants his own hardware. So why waste time messing around with someone else's VM if you just want to test a specific model spec. That is the TL;DR.
Have you tried llama.cpp with unsloth and models suited to it? GLM flash? It seemed to allow more models to be tried soon after they are released. Haven’t tried for long term deployment though, that’s the next step.
Highy anecdotal: I have tried various self-hosted models using both vllm and llama.cpp. I am in a situation where I have access to large amount of memory (~320 GB).
While experimenting with quantization I found that there is a non-trivial tradeoff between quality and memory footprint. Overall my experience follows the reported pattern of "2-bit is mwah, 4-bit half decent and 6-bit required for programming. Still, although MiniMax-m2.7 is useable with the 6-bit quantizations that unsloth provides, it felt like such a breath of fresh air when I used the reference full-size model.
I find it difficult to say why. I had mostly the same setup as before (parsing had to be slightly adjusted in Zed). Aside from not experiencing the thinking loops (where minimax would get stuck generating the same sentences over and over) there is little evidence of any real improvement (although the average thinking time felt shorter).
I would recommend against very low quantizations of GLM 5.0/5.1/5.2 or Kimi 2.5/2.6. Smaller models were more reliable, and therefore more useful.
I only have access to 96GB VRAM locally, but I'd agree with the general approach of avoiding lower quantizations, often anything below Q8 seems to suffer greatly on quality and seemingly never worth going below it, better to go for smaller model in that case.
With the exception of DwarfStar + DS4-Flash with IQ2_XXS quantization, which somehow seems to not suffer as much as I'd thought. I'd still opt for a smaller model + at least Q8.
I have tried llama-cpp, vllm is nicer (ray, handles queueing, doesn't have the cache invalidation bug for qwen/gemma models) and unsloth has toxic employees in their discord.
I've run 2 qwen/gemma @8bit with full context window side-by-side. Right now I have 4 models on my spark (qwen36moe, embedding, reranker, qwen3-1.7B) to support my markdown kb tool.
The setup is not as capable, but still good and gets better with models/algos. To me, it's more about the freedom to tinker, freedom from token bill anxiety, and potential right to compute should the government/oligarchy decides it gets to decide who can access which models.
> unsloth has toxic employees in their discord
Would you mind elaborating on this?
Sure,
I shared a project in their #research channel where I used their qwen36moe quant to refresh my PhD research. The channel had a topic that ended with something like "and all things research..."
One of their people accused me of self-promotion, and I reiterated that I shared it in that channel because it was their quant doing something (I thought) interesting as a research model. The number of people interested in the topic can be counted on your hands (in binary).
They remained accusatory, made it personal, and then started deleting messages. I suppose I escalated a bit (from their perspective), saying how this was not a good first encounter, they could have asked me to move it instead of just deleting it. Then they deleted every message, including all of their own, and put me in timeout. Erased from history, unable to participate, and so I left.
A coworker of mine (ML guy) is also sus about their quants, not nefarious, more that their benchmark results do not mean they are better, possibly skewed / benchmaxxed.
How about Qwen3.6? What sort of prefill/decode rates?
Edit: 3.6 not 3.7!
So far there aren't any open weight model releases for the Qwen 3.7 family.
> So far
Someone's optimistic
I'm hoping the decision makers at Qwen notice how influential the 3.6 series is while the 3.7 series has had very little attention at all.
(Of course for all I know the 3.7 series is doing incredibly well in China, but I've seen almost no buzz around it from the circles that I inhabit.)
My impression is that with the latest round of high-profile releases, the open-weight "market" is coalescing around two players, DS4 Flash for speed and GLM 5.2 for smarts. Qwen is being left behind to pick up the scraps for the terminally GPU-poor.
We know they have what it takes to fight back, and they know it... so I agree, there's no reason not be optimistic about future Qwen releases. But then I've never really understood what motivates these releases in the first place.
Qwen still have the best models that actually run on a laptop - Gemma 4 is their best competition there.
2 LLMs at the same time? I've always wanted to do that
Author here. Quick context the post doesn't quite spell out:
The tool_choice="auto" failure on Qwen3-Next isn't a parser issue — the model reasons inside <think>, decides, and never emits the tool call. No error, just empty tool_calls. The fix was swapping the backbone from Thinking to Instruct, not tuning any parser flag.
The "load the bigger model first, size the smaller against actual residency" playbook generalizes to anything with shared CUDA framework overhead. The ~5 GiB framework floor shows up even at small gpu_memory_utilization values — plan against actuals, not targets.
From the Codex system prompt (verbatim):
```
(...) - Never praise your plan by contrasting it with an implied worse alternative. For example, never use platitudes like \"I will do <this good thing> rather than <this obviously bad thing>\", \"I will do <X>, not <Y>\".
- Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query. (...)
```
It seems the OpenAI people added that first bullet to specifically address the tendency the model has, as seen in the parent comment. The goblin stuff coincidentally appears right after in the system prompt, so in included it as a bonus.
FWIW the mere fact that the goblin stuff is necessary and is in the system prompt suggests to me that OpenAI's approach of training ridiculously large models that can do everything for everyone is hopelessly cooked.
Though I concede it is not that much different than straightening the tie of your most valuable employee before you unwisely put them in front of a client and saying "please don't tell them about the regressions they didn't notice and remember, they don't want things explained in allegories drawn from the Silmarillion".
Can you try and tune your Claude or whatever LLM you're using for your text to phrase things in plain English. Way less use of antithesis, at least. You can probably find a skill for it, if not get an LLM to write your own.
The article was clearly written by an LLM. Please say so at the top.