Still needs 32-64GB memory to run it locally. 64GB Macbook pro with an M5 chip costs more than 4k Euros in Germany. A more practical model would be a language specific (e.g Python or JVM language) and excellent at tool calling and reasoning. Maybe that way they can shrink it even more.
I'm sooo happy I pulled the plug on upgrading and getting a new laptop (with 64 GB RAM) last summer. Feels like it was just in time before the exponential price jumps.
I would love to see any good research projects about it but i have the feeling that Frontier with MoE is making too fast of a progress so that a customized model would always be worse and that the MoE part is actually going somehow in this direction.
On the other hand, at the GTC was a talk about coding in different lanugage (like spanish) and explaining that the quality between spanish and english is relevant different.
But i have not found a good article about the impact of learning data with practical experiments or even if the order of the learning data matters.
At least I think i remember that Meta mentioned having better and less data can be better than more data with lower quality.
As long as these models can explain to you facts about any other topics, its still overfitted for the task though.
I feel like we’ve had this discussion before. From what I remember, specialized models rarely do that much better than general ones, hence no mode Codex models.
Even if you had a 64GB machine: Are you willing to reserve 90% of your memory to run a LLM? With dirt cheap models like deepseek-v4-flash that will run "forever" on $10, the answer for me is clearly: no.
I'm waiting for the speed/quality per dollar metric to go down a little bit further and then I will def run it at home.
Its not just that you send a sentence to an API endpoint, you always send EVERYTHING to that agent as a context.
You want to analyse your spending history? You now send everything to someone.
Either no one cares but understands this implication on how easy it is to really capture you or no one really things about it.
But i'm a lot more diligent on what I send. I disabled the gemini activity feature for example because google started telling me that my stuff could be reviwed by humans.
I see many people saying deepseek and other chinese providers have always been profitable. Also they show their training costs publicly. Can't say for sure since I have not used it personally, but I think they'll for sure outlive the western SOTAs.
Deepseek flash is open weight, this means we can download and run that model without any connection to deepseek, no data/tokens/usage data ever reaches them. They cannot make us their product.
Meta did not abandon opensource. I would love to see a smaller distill, or a moe of this size but the benchmarks seems competetive as long as it isnt benchmaxed witch i would not be suprosed if it is.
Optimizing speed is really the way to go.
Yet 24GB is not what everyone can afford.
Maybe we could take some of those 56tk/s and transfer into some free RAM space using MoE loading ? I'd be glad with a less than 10GB and more than 6tk/s model.
Unfortunately this is just the entry price for LLMs. With the exception of the Qwen 27B models, I personally haven’t found a ton of use cases for models less than 200B. With the right setup, fine tuning, etc, you can make small models do cool things, but hard to please everyone given the insane hardware costs at the moment and the comparably cheap API costs.
What I think would be perfect is a model that could run on a single DGX spark and be competitive with DSV4 Flash 731. Flash is already a game changer. Hopefully meta plans on this, like the old 70b. V4 flash is smart enough for any use but slightly too big. 27b-30b isn’t intelligent enough.
This model I think will be too slow for that on Spark, even at 4 bit quant.
It's a dense model, not MoE like e.g. Qwen 35b or Gemma 4 26B A4B. On a Spark it will be memory bandwidth limited
I haven't tried yet (working on it) but back of the napkin estimate puts it at around 15tok/s even after converting to NVFP4. Prefill would be much higher though. That 15tok/sec is pretty typical for dense models of this size:
NVFP4 Q/K/V/O and MLP projections: ~13 GB/token
BF16 attention gates: ~3 GB/token
BF16 LM head: ~2.5 GB/token
Total: ~18.9 GB/token
At 273 GB/s, that gives a bandwidth-only ceiling of about 14.5 tok/s; actual performance would be lower.
If there is anything meta can do to regain hearts other than owning up their evil deeds, radically change their business model and paying up for taxes and damages, then the world is truly fucked and corporations will continue to win.
The least they could do, after ruthlessly bombarding my employer's servers with requests, ignoring the robots.txt, scraping everything, and incurring significant Google Maps costs for us in the process.
I like this class of model. Multi-token prediction makes it viable to run dense models at not-too-far-off speeds as MoE models with much better intelligence.
The submission’s title (open weights 30B local coding model) is luckily wrong: This is meant to be a general agentic model.
It even comes pre-quantized and with a MTP/drafter model. Looking good!
Let’s hope they aren’t dishonest with the benchmarks this time …
The more open weight models get released the greater the market for personal and small business oriented hardware to run these models. This will drive lower cost hardware, which has stagnated in recent years due to most software not needing the performance and capacity.
The memory math is the part I keep re-reading. 4-bit gets the LM under 20GB, but they're explicitly budgeting the KV cache, the perception encoder, and the DFlash drafter into the same 24GB envelope. On a 4090 that leaves maybe 3-4GB of KV once the drafter and encoder are resident, so long agent traces are going to spill or truncate — which is exactly the workload this model is trained for. I'd like to see the K-Quant-17GB numbers reported at 64k+ context, not just conversational length.
"Minimal to no degradation on agentic tasks" from quantization is also a strong claim. In my experience 4-bit shows up first in tool-call schema adherence — malformed JSON args, wrong enum values — before it moves benchmark averages. Does the report break down tau-Bench / MCP-Atlas per quant level?
Will be interesting to see how Qwen3.8 27B compares against this once it releases this week. Seems like dense 30B is back in fashion?
EDIT: An open weight version of Muse Spark 1.2 is going to be released as well:
https://x.com/alexandr_wang/status/2086756152034066792
https://xcancel.com/alexandr_wang/status/2086756152034066792
Makes me feel hopeful. Things felt more positive around the llama 3 era. Now it’s like a dark, dreadful race.
[delayed]
It’s really interesting timing, Qwen over thinking is what kills it for me. I’m just glad we have more options in this size class now.
Still needs 32-64GB memory to run it locally. 64GB Macbook pro with an M5 chip costs more than 4k Euros in Germany. A more practical model would be a language specific (e.g Python or JVM language) and excellent at tool calling and reasoning. Maybe that way they can shrink it even more.
I'm sooo happy I pulled the plug on upgrading and getting a new laptop (with 64 GB RAM) last summer. Feels like it was just in time before the exponential price jumps.
Bought an M1 64 GB for 2000 euro’s second hand a year ago. That was sweet
Pulled the trigger?
There have been discussions on language specific not really being a relevant change to reduce size.
I would love to see any good research projects about it but i have the feeling that Frontier with MoE is making too fast of a progress so that a customized model would always be worse and that the MoE part is actually going somehow in this direction.
On the other hand, at the GTC was a talk about coding in different lanugage (like spanish) and explaining that the quality between spanish and english is relevant different.
But i have not found a good article about the impact of learning data with practical experiments or even if the order of the learning data matters.
At least I think i remember that Meta mentioned having better and less data can be better than more data with lower quality.
As long as these models can explain to you facts about any other topics, its still overfitted for the task though.
I feel like we’ve had this discussion before. From what I remember, specialized models rarely do that much better than general ones, hence no mode Codex models.
Even if you had a 64GB machine: Are you willing to reserve 90% of your memory to run a LLM? With dirt cheap models like deepseek-v4-flash that will run "forever" on $10, the answer for me is clearly: no.
I'm waiting for the speed/quality per dollar metric to go down a little bit further and then I will def run it at home.
Its not just that you send a sentence to an API endpoint, you always send EVERYTHING to that agent as a context.
You want to analyse your spending history? You now send everything to someone.
Either no one cares but understands this implication on how easy it is to really capture you or no one really things about it.
But i'm a lot more diligent on what I send. I disabled the gemini activity feature for example because google started telling me that my stuff could be reviwed by humans.
[delayed]
"With dirt cheap models like deepseek-v4-flash that will run "forever" on $10, the answer for me is clearly: no."
When it's free, you are the product.
I see many people saying deepseek and other chinese providers have always been profitable. Also they show their training costs publicly. Can't say for sure since I have not used it personally, but I think they'll for sure outlive the western SOTAs.
Deepseek flash is open weight, this means we can download and run that model without any connection to deepseek, no data/tokens/usage data ever reaches them. They cannot make us their product.
Meta did not abandon opensource. I would love to see a smaller distill, or a moe of this size but the benchmarks seems competetive as long as it isnt benchmaxed witch i would not be suprosed if it is.
Optimizing speed is really the way to go. Yet 24GB is not what everyone can afford. Maybe we could take some of those 56tk/s and transfer into some free RAM space using MoE loading ? I'd be glad with a less than 10GB and more than 6tk/s model.
Unfortunately this is just the entry price for LLMs. With the exception of the Qwen 27B models, I personally haven’t found a ton of use cases for models less than 200B. With the right setup, fine tuning, etc, you can make small models do cool things, but hard to please everyone given the insane hardware costs at the moment and the comparably cheap API costs.
I don't thinnk just MoE will solve it. If you hit constantly different expert layers, you can't outsource layers efficently and have to swap it in.
MoE will be faster because it will read less memory for sure, you still have to have it though.
What I think would be perfect is a model that could run on a single DGX spark and be competitive with DSV4 Flash 731. Flash is already a game changer. Hopefully meta plans on this, like the old 70b. V4 flash is smart enough for any use but slightly too big. 27b-30b isn’t intelligent enough.
DSV4 Flash 0731 already runs on RTX 4090 24GB + 128GB system RAM at a usable tok/s and quantization.
You personally? Just curious. Context window is also a factor and ram isn’t really cheap. Sparks are assembled units which I like.
This model I think will be too slow for that on Spark, even at 4 bit quant.
It's a dense model, not MoE like e.g. Qwen 35b or Gemma 4 26B A4B. On a Spark it will be memory bandwidth limited
I haven't tried yet (working on it) but back of the napkin estimate puts it at around 15tok/s even after converting to NVFP4. Prefill would be much higher though. That 15tok/sec is pretty typical for dense models of this size:
NVFP4 Q/K/V/O and MLP projections: ~13 GB/token
BF16 attention gates: ~3 GB/token
BF16 LM head: ~2.5 GB/token
Total: ~18.9 GB/token
At 273 GB/s, that gives a bandwidth-only ceiling of about 14.5 tok/s; actual performance would be lower.
Meta knows how to win back developer's hearts .... let's see if they have the goods
If there is anything meta can do to regain hearts other than owning up their evil deeds, radically change their business model and paying up for taxes and damages, then the world is truly fucked and corporations will continue to win.
good to see new open weights releases from meta
The least they could do, after ruthlessly bombarding my employer's servers with requests, ignoring the robots.txt, scraping everything, and incurring significant Google Maps costs for us in the process.
good looking showing too, which is excellent.
Looking forward to giving this a try with llama.cpp. I’m watching the open-weights competition with high expectations.
The favourable comparisons to Gemma 4 and qwen3.6 look promising!
Wow, Meta is back (at least for now)!
I like this class of model. Multi-token prediction makes it viable to run dense models at not-too-far-off speeds as MoE models with much better intelligence.
The submission’s title (open weights 30B local coding model) is luckily wrong: This is meant to be a general agentic model.
It even comes pre-quantized and with a MTP/drafter model. Looking good!
Let’s hope they aren’t dishonest with the benchmarks this time …
The more open weight models get released the greater the market for personal and small business oriented hardware to run these models. This will drive lower cost hardware, which has stagnated in recent years due to most software not needing the performance and capacity.
The memory math is the part I keep re-reading. 4-bit gets the LM under 20GB, but they're explicitly budgeting the KV cache, the perception encoder, and the DFlash drafter into the same 24GB envelope. On a 4090 that leaves maybe 3-4GB of KV once the drafter and encoder are resident, so long agent traces are going to spill or truncate — which is exactly the workload this model is trained for. I'd like to see the K-Quant-17GB numbers reported at 64k+ context, not just conversational length.
"Minimal to no degradation on agentic tasks" from quantization is also a strong claim. In my experience 4-bit shows up first in tool-call schema adherence — malformed JSON args, wrong enum values — before it moves benchmark averages. Does the report break down tau-Bench / MCP-Atlas per quant level?