I lament the comments saying this in any way redeems Meta (the company).
The researchers releasing this stuff have almost nothing to do with Meta other than being bankrolled by the slaughterhouse.
You aren't the customer, you are the pawn in big tech's game of thrones. Your good will is a commodity to be traded, almost literally. It will be used against you the moment it's convenient. This is open weights because Meta couldn't monetize it in any other way than to cloud developer's judgement of their reputation.
But I guess most people just don't care.
I'm glad it's open. It does not make me think any better of Meta.
Meta can never be redeemed, but it's still valid to admit that FB at one point had a very badass engineering culture.
They're one of 2 companies I would absolutely never work for (weapons etc aside). FB's recruiters hounded me so often I requested that they blackball me. The day they became Meta, I learned this by checking my email to see that they started trying to reach out again. I once again requested that they blackball me. This by extention taints OAI, the other company I'll never work for.
After a few hours with Glimmer I'm pretty impressed. It's better than the benchmark scores seem to indicate compared to Qwen 3.6 27B. I'm very excited for 3.8
Muse Glimmer doesn't redeem Meta, but it's a contribution to the commons and the Apache 2.0 licensing is an improvement from the restricted licenses attached to Llama. If even Meta can use a permissive license for its model weights, so can any other company.
To be honest the main issue with meta has never been around open/closed software. They've also done react, Cassandra and some other bits. But this, like their open weights is like a feather pressing down on the scale compared to things like promoting genocide in Myanmar, enabling Cambridge analytica, creating a huge closed ecosystem which dominate(s/d) local community communication, mandating doxxed communication, trying to replace actual community communication with algorithmic nonsense etc.
Meta and its products, as a whole, is a threat to your kids, your mental health, your community's health and the planet as a whole. It is just sad and very repulsive everyone fell so easily addicted to their social drug. Yes - it is a drug, and it is hard to get off from.
Nothing redeems them at this point of time, they are doing exactly ZERO to redeem. Tossing open weight models (not opensource!!) is not a basis for redemption, and does not constitute remorse in any way. Trying to portray it as such is complicity to META's crimes against humanity.
Any retort to do this like “but why would they just openly release this” pretty much answers itself. Public relations.
If a company can spend money to redeem itself then, well, it can (game theoretically or whatever) do whatever it wants in the future and then spend money to wipe the slate clean.
Based on the benchmarks, it seems that Muse Glimmer barely edges out against Qwen3.6 27B, except for tool-calling skills (MCP, etc.). I wouldn't be surprised if they released it now because they are afraid they wouldn't beat Qwen3.8 27B.
Do AI companies make release plans based on upcoming other models like this? I would think all the processes that go into the repository and weight infrastructure pre-training, checkpointing, knowledge distillation, model compression, post training pipeline, ecosystem integrations, inference API, benchmarking, human eval/safety/alignment, docs, etc... all that dictates the release schedule.
Yes, not every model release is reactionary to other labs. Either they had hints for the release of other models or they cut efforts in late stage testing of the models to hit these earlier release dates. There’s always some flexibility. And there’s certainly the incentive to cannibalize the news cycles for competitor models.
the last few items there (benchmarking, human evaluation, docs) can be rushed or skipped by leadership if they want to beat comp. they probably spend a few weeks on those things normally
There has been a long history of AI model releases made shortly before or after a major planned release by another company. Almost always to upstage or steal thunder.
Just recently, Minimax H3 released as open weights on the eve of Seedance 2.5 global availability. It's not as good, but it's good enough and it's completely open.
Flux 3, which is nowhere near as good as either, suddenly announced their release once news of these other two became public. They knew if they waited they'd be ignored. It didn't really help them much, unfortunately.
The LLM releases are even more rivalrous.
And don't forget all of the competing launches planned before Google IO or major release events.
Companies like to eat into the news and press cycle of their rivals.
I've seen it here on HN (it's particularly noticeable via the /active page) multiple times. If Google, OpenAI or Anthropic release something significant, odds are good you'll see a headline from one of the others.
If you start counting since WaveNet or BERT, it's been ages. Especially when it feels like decades of advancements happen every single year, and rival labs are always trying to one up each other.
You're mixing up sizes of different quants. The 60GB is unquantized, and Qwen's unquantized size is around 54GB. Their sizes as like quantization levels are similar.
Considering how all the big players are playing fast and loose with quotas, billing and undisclosed testing that burns your tokens even faster, it can't happen soon enough.
It is interesting but it does look like a careful distillation of (Spark and) biggers open-weight models.
The progress compared to Qwen3.6 27B is good, not that impressive, it's a 4 months old model. (kuto to them to compare to 27B dense and not 35B MoE, it's more fair to do so).
It is very probable that Qwen3.8 27B will crush Glimmer-30B on most benchmarks.
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 trigger 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.
That commenter you're replying to knows that. The original commenter before them wrote "pulled the plug" which is different and doesn't quite apply here (actually implies the opposite of what they meant to say).
I think if there's going to be advantages to making smaller, more targeted models, those advantages will probably come from targeting specific domains, not from targeting specific languages.
I think that if an LLM can't abstract over the differences between Python and C++, it probably will have an even harder time abstracting over the differences between writing code that manages a webserver, and writing code that does aerodynamic simulations.
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.
Capability in LLM's is distributed throughout the manifold in subspaces. Even worse, the subspaces exist in superposition.
That is to say, there is no single 'python' part of the model. The python bit is spread throughout the entire model and overlaps with other pieces that have similar, but unrelated, capabilities. For example the python subpspace might be partially in superposition with cupcake recipes, Esperanto, and calculus. We need calculus in a coding agent but not the other two. However, separating them cleanly is almost impossible, and even identifying them is tough.
Internally the manifolds are highly inefficient and nothing like you would imagine something humans built would be designed. It's more like something that evolved in nature.
With MOE you train a router designed to select which parts to activate. The router itself is a trained neural network and the 'experts' are usually not really things like 'python'. They're just the functional subspaces I described above.
Again, those subspaces are all somehow inextricably correlated and live in complex superposition spread throughout the manifold. The router doesn't know (or care) WHY those sections get lit up it just learns which ones to activate to optimize it's own reward function. So maybe it learns to activate "logic", "python" and "cupcake recipes in esperanto" whenever it see's something that kind of looks like python. It's not the best answer, it's just the best answer the tiny router could figure out.
It's all wildly complicated and inefficient, and works nothing like any reasonable human would imagine that it SHOULD operate.
There was some paper about routing at training bio-knowledge into a particular region of the model, which you then can cutoff when serving. But you probably lose some efficiency since maybe you sized that region too small/too big.
If you can’t do it cheaper on your own hardware it does make you wonder how much of the cost of inference those large LLM providers are eating? Datacenter hardware isn’t magic.
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.
Yeah, it does feel a bit silly with my encrypted disks, encrypted backups, unique passwords, advanced router, etc, while I send everything I do in plain text to anthropic.
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.
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.
OpenAI apparently runs a profitable inference business with 40% gross margin, but their advertising budget is nutso and their real costs are pretraining and research. I suspect Deepseek's comp is not predicated on capturing the lightcone of all future value, some googling insinuates their top pay is $212K US which would support that suspicion. Compare and contrast with the $1.35M and up at OpenAI.
Ah yes, I'm sure Trovalds and Stallman are harvesting my data through free software, aren't they? This argument is used by boomers who were fed cold war era propoganda that surely everybody is selfish, and you're always at fault.
I don't understand the desire to run own AI models for programming locally. No laptop is ever going to be as powerful and energy efficient to run anything close to OpenAI, Anthropic or Google models. A model you can run on a loptop is simply not going to work as well as it's needed for programming. Small models for linguistic work fine, but anything more sophisticated simply won't provide enough resources or power. Or models would need to be significantly dumbed down - then why use them at all? So far the idea of carrying a "thin" or "thin"-like device looks more reasonable to me, while running AI on your own server.
> A model you can run on a loptop is simply not going to work as well as it's needed for programming
The models you can run on a high-spec laptop today are approximately where frontier models were 12-18mo ago (albeit at a lower tok/s rate). If you scan back through hn comments from that era, you’ll find plenty of people saying “this is powerful enough to massively increase my productivity”.
Not always! I get 80-100 tok/s from Qwen 3.6 35B-A3B on a MacBook Pro thanks to MTP. With long contexts that dips to around 50-60. However, prefill is much slower than API models. So it becomes really, really, really critical to not have cache misses.
I’m quite optimistic about the long-term future of local LLMs for privacy and cost control reasons. An LLM running on my own hardware, even if it’s not a laptop but a home server, is one where I don’t need to worry about token limits, token fees, privacy, and “rug-pulling” from the vendor.
In the short term, the big challenge is being able to afford hardware that can run a ~30B model. Last month I got to experiment with LLMs on a NVIDIA RTX 6000 Ada Generation as a visiting researcher during my summer break. I see the power of local LLMs for agentic coding; they’re no Claude, but they are quite useful. I wish I had gotten into local LLMs before hardware has gotten prohibitively expensive and in some cases unavailable; Apple discontinued certain Mac Minis and Mac Studios with high amounts of RAM due to the RAM shortage.
Hopefully high RAM prices don’t become a new normal, though the next year or two doesn’t look good.
For some companies there might be a need to run them locally. For instance, Apple decided to run LLMs on the phone locally. I guess it depends on how important latency and privacy are. Perhaps Meta is looking at how much interest for those local models is there.
I've never done it but would be interested because it cuts out the burden of worrying about costs. Maybe I'm mistaken on energy cost here. There's a constant raincloud that follows me around regarding limits, and it would be nice to shake that.
I've been able to accomplish incredible feats (for myself) since GPT-4, so model intelligence is secondary.
Some interesting findings from the chat template designs:
1. The template name is Onyx ATEM as found in the tool call exception message
2. It appears to be following a harmony-style chat template. But the tool use seems to be a xml like :<atem:function_calls> / <atem:invoke> / <atem:parameter>
The XML tags are similar to <antml:xxx>, which is obviously Anthropic ML (or ANTrophic xML).
I think it’s likely 3; meta in reverse. While tokenisers and preprocessing can catch it, you want your special tokens to be unique and not present in the original corpus. <meta: is likely too common.
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.
For the same price as a DGX Spark here (A$8499) I can buy roughly 544GB of DDR5-5200MHz from retail; which on a quad channel platform would deliver ~160gb/s real world; and ~320gb/s with octa channels (Xeon, Threadripper Pro).
If you can afford it or somehow find a used unit, you can go Epyc for 12 channels.
8/12 channel DDR5 will beat DGX Spark in inference/decode even without a GPU of any kind, as it’s memory bandwidth bound, and the Spark tops out at ~240gb/s real world.
You are paying an extraordinary amount of money for the convenience of a super small unit, with “better than Aliexpress no brand” support and documentation, but still early software support and likely a trip to the community forums for most tasks.
Please don’t pay another 70-80% gross margins on top of already inflated DRAM prices unless you want to nonetheless. The Spark IS really nice if you want to test out ConnectX or if you really need something small and compact and quiet.
And used Adas or even Ampere NVIDIA workstation GPUs can come with a lot of VRAM and be “reasonable”.
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.
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.
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.
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.
I don't think outside of the Big 3 (Ant, OAI, GDM), given the strong competition from China, any other Lab has a chance at capturing the coding market if they aren't open weights (save for xAI whose latest Grok looks every bit good & will probably rely on Cursor for distribution instead of going open weights). There's literally no other selling point, as the capabilities have mostly converged by now among the chasing pack.
There was a good discussion yesterday on the DeepSeek Flash release thread about this.
There's a large market, very large, who want the best regardless of what it costs. Probably a large enough market to keep that domain of research afloat (as opposed to shifting research manpower to cost cutting).
The reasoning is just that the marginal cost of AI is very secondary to fixed costs of the businesses themselves; it's not an excuse to sacrifice performance.
Those two offer MoE variants, this doesn't seem to.
Dense model makes it dog slow on anything without HBM. Max 15tok/sec on decode on DDR5 systems like a Spark or a Strix Halo -- and that's at 4 bit quant.
Dense models run at a very usable speed (Qwen 3.6 was running at ~50t/s last I looked) on my dual 7900 XTX desktop. (And before anyone brings it up, I did not buy them for this purpose, so the up-front cost is irrelevant in my case.)
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 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.
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 …
As an industry, I wish we would stop calling these things "open weight" because it is too easy to confuse with actual "open source", which they are not.
Photoshop source code+ OSI license = open source
Photoshop binary you can run on your own computer = open weight
Photoshop SaaS web app = closed, proprietary (Opus, GPT, etc.)
"Open weight" models are still just binary blobs that are completely inscrutable. It's like bringing home a dog from the rescue and just hoping that it doesn't have a tendency to bite kids in the face. You just can't know. The only thing that you can do is try to add more training (fine tuning) telling it not to bite kids.
I don't think the FOSS community has ever accepted this, but somehow we're feeling like it is okay now.
Photoshop binary you can run on your own computer = open weight
I don't think this is a correct analogy. You are not allowed to distribute modified versions of the Photoshop binary. Most open weight model licenses allow you to make and distribute your own finetunes, etc.
I believe that comparing LLMs with traditional deterministic software is fundamentally misleading. It is extremely difficult to truly interpret what LLMs do internally, and as of now, nobody fully understands it. Even if you trained the LLM yourself, there is no source code you can simply read and learn from.
Sure, having information about how these models were trained is helpful for reproducibility, but it is basically impossible for anyone without substantial capital and access to the same (likely copyrighted) data to reproduce the model. For normal users, owning the model weights essentially means owning 100% of the model, you can inspect and study the weights in much the same way as the lab that produced the model can, you can modify the weights, and you can use and distribute them if the license allows you to
Given an open weights model trained to sometimes bite kids, we can’t train it to not bite kids, even though billions of dollars of research have been thrown at this open problem.
Given an open weights model trained to never bite kids, you can get it to bite kids with 10 prompts and a linear projection, the known simple algorithm doesn’t even need a backwards pass.
It is useful to indicate you can run the weights on your own hardware. That’s categorically different from most other commercial offerings. It’s as if your adobe example ignores the reality that would exist had photoshop been invented in 2019: cloud only.
I am extremely well aware of how rescues evaluate dogs. And I'm also fully aware that they do not know the full history of the dog. They go through a limited set of testing and interrogation to evaluate the safety of the dog. That's it.
I lament the comments saying this in any way redeems Meta (the company).
The researchers releasing this stuff have almost nothing to do with Meta other than being bankrolled by the slaughterhouse.
You aren't the customer, you are the pawn in big tech's game of thrones. Your good will is a commodity to be traded, almost literally. It will be used against you the moment it's convenient. This is open weights because Meta couldn't monetize it in any other way than to cloud developer's judgement of their reputation.
But I guess most people just don't care.
I'm glad it's open. It does not make me think any better of Meta.
Meta can never be redeemed, but it's still valid to admit that FB at one point had a very badass engineering culture.
They're one of 2 companies I would absolutely never work for (weapons etc aside). FB's recruiters hounded me so often I requested that they blackball me. The day they became Meta, I learned this by checking my email to see that they started trying to reach out again. I once again requested that they blackball me. This by extention taints OAI, the other company I'll never work for.
After a few hours with Glimmer I'm pretty impressed. It's better than the benchmark scores seem to indicate compared to Qwen 3.6 27B. I'm very excited for 3.8
There is literally not a single comment like this, the only off topic comment like this is yours.
How is this non-sequitor the top comment?
Muse Glimmer doesn't redeem Meta, but it's a contribution to the commons and the Apache 2.0 licensing is an improvement from the restricted licenses attached to Llama. If even Meta can use a permissive license for its model weights, so can any other company.
To be honest the main issue with meta has never been around open/closed software. They've also done react, Cassandra and some other bits. But this, like their open weights is like a feather pressing down on the scale compared to things like promoting genocide in Myanmar, enabling Cambridge analytica, creating a huge closed ecosystem which dominate(s/d) local community communication, mandating doxxed communication, trying to replace actual community communication with algorithmic nonsense etc.
Meta and its products, as a whole, is a threat to your kids, your mental health, your community's health and the planet as a whole. It is just sad and very repulsive everyone fell so easily addicted to their social drug. Yes - it is a drug, and it is hard to get off from.
Nothing redeems them at this point of time, they are doing exactly ZERO to redeem. Tossing open weight models (not opensource!!) is not a basis for redemption, and does not constitute remorse in any way. Trying to portray it as such is complicity to META's crimes against humanity.
I think it’s also worth pointing out that that there are numerous less evil options to choose from.
Perhaps none of the AI companies are shining examples of high ethics, but basically all of them have ethical high ground over Meta.
At least Anthropic isn’t sending private videos from pervert glasses to contract workers in Africa. It’s a low bar but it’s a bar nonetheless.
Any retort to do this like “but why would they just openly release this” pretty much answers itself. Public relations.
If a company can spend money to redeem itself then, well, it can (game theoretically or whatever) do whatever it wants in the future and then spend money to wipe the slate clean.
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
Based on the benchmarks, it seems that Muse Glimmer barely edges out against Qwen3.6 27B, except for tool-calling skills (MCP, etc.). I wouldn't be surprised if they released it now because they are afraid they wouldn't beat Qwen3.8 27B.
Do AI companies make release plans based on upcoming other models like this? I would think all the processes that go into the repository and weight infrastructure pre-training, checkpointing, knowledge distillation, model compression, post training pipeline, ecosystem integrations, inference API, benchmarking, human eval/safety/alignment, docs, etc... all that dictates the release schedule.
Yes, not every model release is reactionary to other labs. Either they had hints for the release of other models or they cut efforts in late stage testing of the models to hit these earlier release dates. There’s always some flexibility. And there’s certainly the incentive to cannibalize the news cycles for competitor models.
the last few items there (benchmarking, human evaluation, docs) can be rushed or skipped by leadership if they want to beat comp. they probably spend a few weeks on those things normally
There has been a long history of AI model releases made shortly before or after a major planned release by another company. Almost always to upstage or steal thunder.
Just recently, Minimax H3 released as open weights on the eve of Seedance 2.5 global availability. It's not as good, but it's good enough and it's completely open.
Flux 3, which is nowhere near as good as either, suddenly announced their release once news of these other two became public. They knew if they waited they'd be ignored. It didn't really help them much, unfortunately.
The LLM releases are even more rivalrous.
And don't forget all of the competing launches planned before Google IO or major release events.
Companies like to eat into the news and press cycle of their rivals.
I've seen it here on HN (it's particularly noticeable via the /active page) multiple times. If Google, OpenAI or Anthropic release something significant, odds are good you'll see a headline from one of the others.
>long history
Seems a bit premature of a statement lol
If you start counting since WaveNet or BERT, it's been ages. Especially when it feels like decades of advancements happen every single year, and rival labs are always trying to one up each other.
Yes, and also waiting for the next iteration of Gemma. Muse or Qwen are optimized for coding, while IMO Gemma is still better for non-coding tasks.
https://x.com/osanseviero/status/2086107547535122767
Makes me feel hopeful. Things felt more positive around the llama 3 era. Now it’s like a dark, dreadful race.
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.
yes i think everyone is waiting to see that ;d, i've been on qwen for the last year and a half now.
> Seems like dense 30B is back in fashion?
Surprising that Meta don't host this model, even as rate-limited free-tier.
> open weight version of Muse Spark 1.2
Wait. Is this "version" different from what Meta serves?
It seems worse than 3.6, but a bit smaller.
UPD. was wrong on smaller, it's actually much larger
How is 30B smaller than 27B?
It uses fractal compression
They say it is trained with quantization awareness, so it should only be 15GB or so. Qwen was only trained in FP8 with QAT.
UPD, NVM, got misled by comments here. It is actually almost 60 GB so much larger
You're mixing up sizes of different quants. The 60GB is unquantized, and Qwen's unquantized size is around 54GB. Their sizes as like quantization levels are similar.
With the business model for API based LLMs looking iffy at best it seems like we’re heading back to the “server under your desk” era of IT again.
Considering how all the big players are playing fast and loose with quotas, billing and undisclosed testing that burns your tokens even faster, it can't happen soon enough.
The comparison set is Gemma4-31B and Qwen3.6-27B, not the current Qwen
Fair on size, but the headline numbers are against a model a generation back
It is interesting but it does look like a careful distillation of (Spark and) biggers open-weight models.
The progress compared to Qwen3.6 27B is good, not that impressive, it's a 4 months old model. (kuto to them to compare to 27B dense and not 35B MoE, it's more fair to do so). It is very probable that Qwen3.8 27B will crush Glimmer-30B on most benchmarks.
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.
Practically ~20GB with KV cache
> We quantize weights to ~4-bit, bringing the LM under 20 GB. We validated minimal to no degradation on agentic tasks under compression.
https://www.reddit.com/r/LocalLLaMA/comments/1vkgsum/introdu...
I'm sooo happy I pulled the trigger 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.
Pulled the trigger?
lol, you're right, the brainfart completely changes the meaning.
I corrected it.
Common phrase.
Parent used "pulled the plug", are you saying it's applicable here and not "pulled the trigger" like suggested?
That commenter you're replying to knows that. The original commenter before them wrote "pulled the plug" which is different and doesn't quite apply here (actually implies the opposite of what they meant to say).
Bought an M1 64 GB for 2000 euro’s second hand a year ago. That was sweet
paid 2.7k € for this same build new in Dec 2023, that was also sweet (still is)
I think if there's going to be advantages to making smaller, more targeted models, those advantages will probably come from targeting specific domains, not from targeting specific languages.
I think that if an LLM can't abstract over the differences between Python and C++, it probably will have an even harder time abstracting over the differences between writing code that manages a webserver, and writing code that does aerodynamic simulations.
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.
Capability in LLM's is distributed throughout the manifold in subspaces. Even worse, the subspaces exist in superposition.
That is to say, there is no single 'python' part of the model. The python bit is spread throughout the entire model and overlaps with other pieces that have similar, but unrelated, capabilities. For example the python subpspace might be partially in superposition with cupcake recipes, Esperanto, and calculus. We need calculus in a coding agent but not the other two. However, separating them cleanly is almost impossible, and even identifying them is tough.
Internally the manifolds are highly inefficient and nothing like you would imagine something humans built would be designed. It's more like something that evolved in nature.
My current image from a MoE is that the base/core might be the more generic thing and that things like python are part of one expert though.
With MOE you train a router designed to select which parts to activate. The router itself is a trained neural network and the 'experts' are usually not really things like 'python'. They're just the functional subspaces I described above.
Again, those subspaces are all somehow inextricably correlated and live in complex superposition spread throughout the manifold. The router doesn't know (or care) WHY those sections get lit up it just learns which ones to activate to optimize it's own reward function. So maybe it learns to activate "logic", "python" and "cupcake recipes in esperanto" whenever it see's something that kind of looks like python. It's not the best answer, it's just the best answer the tiny router could figure out.
It's all wildly complicated and inefficient, and works nothing like any reasonable human would imagine that it SHOULD operate.
There was some paper about routing at training bio-knowledge into a particular region of the model, which you then can cutoff when serving. But you probably lose some efficiency since maybe you sized that region too small/too big.
That's a very clever approach. Any idea about the papers title or authors? I'd love to look it up.
Well if you're spending thousands on API tokens already, you could just drop the same amount on a 128GB MacBook Pro and that's a one time cost.
If you're dropping thousands on API tokens, you're going to be slowed down at least 10x trying to do everything on a single MBP.
Don't forget about energy usage, you'll probably never break even vs same model on openrouter.
If you can’t do it cheaper on your own hardware it does make you wonder how much of the cost of inference those large LLM providers are eating? Datacenter hardware isn’t magic.
Datacenter hardware can batch at large scale, probably over 90% more energy efficient per token than a MacBook.
> 64GB Macbook pro with an M5 chip costs more than 4k Euros in Germany
Sure, if you want the latest and almost* greatest. You can pick up an M1 Max 64GB for ~1k.
* I guess 128GB also exists
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.
4K bucks buys you around 180 months of <insert AI subscription here> with zero upfront cost.
Problem is that might go away or get nerfed.
then you switch provider, it's not a monopoly
Haha wow. I’m trying to even imagine the AI landscape in 15 years and I can’t.
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.
Yeah, it does feel a bit silly with my encrypted disks, encrypted backups, unique passwords, advanced router, etc, while I send everything I do in plain text to anthropic.
"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.
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.
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.
OpenAI apparently runs a profitable inference business with 40% gross margin, but their advertising budget is nutso and their real costs are pretraining and research. I suspect Deepseek's comp is not predicated on capturing the lightcone of all future value, some googling insinuates their top pay is $212K US which would support that suspicion. Compare and contrast with the $1.35M and up at OpenAI.
Ah yes, I'm sure Trovalds and Stallman are harvesting my data through free software, aren't they? This argument is used by boomers who were fed cold war era propoganda that surely everybody is selfish, and you're always at fault.
It's the size of a big vm. There's nothing wrong with reserving that much working space for one item.
I don't understand the desire to run own AI models for programming locally. No laptop is ever going to be as powerful and energy efficient to run anything close to OpenAI, Anthropic or Google models. A model you can run on a loptop is simply not going to work as well as it's needed for programming. Small models for linguistic work fine, but anything more sophisticated simply won't provide enough resources or power. Or models would need to be significantly dumbed down - then why use them at all? So far the idea of carrying a "thin" or "thin"-like device looks more reasonable to me, while running AI on your own server.
> A model you can run on a loptop is simply not going to work as well as it's needed for programming
The models you can run on a high-spec laptop today are approximately where frontier models were 12-18mo ago (albeit at a lower tok/s rate). If you scan back through hn comments from that era, you’ll find plenty of people saying “this is powerful enough to massively increase my productivity”.
> albeit at a lower tok/s rate
Not always! I get 80-100 tok/s from Qwen 3.6 35B-A3B on a MacBook Pro thanks to MTP. With long contexts that dips to around 50-60. However, prefill is much slower than API models. So it becomes really, really, really critical to not have cache misses.
I’m quite optimistic about the long-term future of local LLMs for privacy and cost control reasons. An LLM running on my own hardware, even if it’s not a laptop but a home server, is one where I don’t need to worry about token limits, token fees, privacy, and “rug-pulling” from the vendor.
In the short term, the big challenge is being able to afford hardware that can run a ~30B model. Last month I got to experiment with LLMs on a NVIDIA RTX 6000 Ada Generation as a visiting researcher during my summer break. I see the power of local LLMs for agentic coding; they’re no Claude, but they are quite useful. I wish I had gotten into local LLMs before hardware has gotten prohibitively expensive and in some cases unavailable; Apple discontinued certain Mac Minis and Mac Studios with high amounts of RAM due to the RAM shortage.
Hopefully high RAM prices don’t become a new normal, though the next year or two doesn’t look good.
For some companies there might be a need to run them locally. For instance, Apple decided to run LLMs on the phone locally. I guess it depends on how important latency and privacy are. Perhaps Meta is looking at how much interest for those local models is there.
I've never done it but would be interested because it cuts out the burden of worrying about costs. Maybe I'm mistaken on energy cost here. There's a constant raincloud that follows me around regarding limits, and it would be nice to shake that.
I've been able to accomplish incredible feats (for myself) since GPT-4, so model intelligence is secondary.
Yet.
Some interesting findings from the chat template designs:
1. The template name is Onyx ATEM as found in the tool call exception message
2. It appears to be following a harmony-style chat template. But the tool use seems to be a xml like :<atem:function_calls> / <atem:invoke> / <atem:parameter>
3. atem: a internal joke of meta in reverse?
https://huggingface.co/meta-models/Muse-Glimmer-30B/blob/mai...
The XML tags are similar to <antml:xxx>, which is obviously Anthropic ML (or ANTrophic xML).
I think it’s likely 3; meta in reverse. While tokenisers and preprocessing can catch it, you want your special tokens to be unique and not present in the original corpus. <meta: is likely too common.
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.
For the same price as a DGX Spark here (A$8499) I can buy roughly 544GB of DDR5-5200MHz from retail; which on a quad channel platform would deliver ~160gb/s real world; and ~320gb/s with octa channels (Xeon, Threadripper Pro).
If you can afford it or somehow find a used unit, you can go Epyc for 12 channels.
8/12 channel DDR5 will beat DGX Spark in inference/decode even without a GPU of any kind, as it’s memory bandwidth bound, and the Spark tops out at ~240gb/s real world.
You are paying an extraordinary amount of money for the convenience of a super small unit, with “better than Aliexpress no brand” support and documentation, but still early software support and likely a trip to the community forums for most tasks.
Please don’t pay another 70-80% gross margins on top of already inflated DRAM prices unless you want to nonetheless. The Spark IS really nice if you want to test out ConnectX or if you really need something small and compact and quiet.
And used Adas or even Ampere NVIDIA workstation GPUs can come with a lot of VRAM and be “reasonable”.
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.
Native dflash support on day 1 helps a lot! High quality speculative decoding speeds up a lot of agentic work.
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.
Gemma4-E4B (4B params) works pretty well as a local wiki, or when you don't have connectivity.
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.
Looking forward to giving this a try with llama.cpp. I’m watching the open-weights competition with high expectations.
good to see new open weights releases from meta
good looking showing too, which is excellent.
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.
Meta released their own 4-bit quant of this model for devices with 24GB VRAM.
That's a modern gaming laptop; cheapest I see in the US with 24GB is $3.5k.
Should be quite a bit faster than the new M5 MacBook Pro, and you can run Linux on it!
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.
> Meta did not abandon opensource
Open weights*
I don't think outside of the Big 3 (Ant, OAI, GDM), given the strong competition from China, any other Lab has a chance at capturing the coding market if they aren't open weights (save for xAI whose latest Grok looks every bit good & will probably rely on Cursor for distribution instead of going open weights). There's literally no other selling point, as the capabilities have mostly converged by now among the chasing pack.
There was a good discussion yesterday on the DeepSeek Flash release thread about this.
There's a large market, very large, who want the best regardless of what it costs. Probably a large enough market to keep that domain of research afloat (as opposed to shifting research manpower to cost cutting).
The reasoning is just that the marginal cost of AI is very secondary to fixed costs of the businesses themselves; it's not an excuse to sacrifice performance.
Meta seems like the one American bigtech that would distill the the other American frontier models. My enemy’s enemy is my friend?
They do distill, their own bigger Muse model.
The favourable comparisons to Gemma 4 and qwen3.6 look promising!
Those two offer MoE variants, this doesn't seem to.
Dense model makes it dog slow on anything without HBM. Max 15tok/sec on decode on DDR5 systems like a Spark or a Strix Halo -- and that's at 4 bit quant.
The benchmark comparison is against the dense variants not MoE
Dense models run at a very usable speed (Qwen 3.6 was running at ~50t/s last I looked) on my dual 7900 XTX desktop. (And before anyone brings it up, I did not buy them for this purpose, so the up-front cost is irrelevant in my case.)
3090/4090 probably would do 40 t/s, for 5090 75 t/s is shown in the blog.
I wish they would release the quantized versions in a safetensor format. Many frameworks can't load PTE and GGUF.
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.
Can I run this on my RTX 5090?
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.
Higher demand for 5090's did not make them cheaper, because Nvidia got much higher margin products to focus on.
The opposite happening because foundries are full to capacity making higher margin stuff.
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 …
As an industry, I wish we would stop calling these things "open weight" because it is too easy to confuse with actual "open source", which they are not.
Photoshop source code+ OSI license = open source
Photoshop binary you can run on your own computer = open weight
Photoshop SaaS web app = closed, proprietary (Opus, GPT, etc.)
"Open weight" models are still just binary blobs that are completely inscrutable. It's like bringing home a dog from the rescue and just hoping that it doesn't have a tendency to bite kids in the face. You just can't know. The only thing that you can do is try to add more training (fine tuning) telling it not to bite kids.
I don't think the FOSS community has ever accepted this, but somehow we're feeling like it is okay now.
Photoshop source code+ OSI license = open source
Photoshop binary you can run on your own computer = open weight
I don't think this is a correct analogy. You are not allowed to distribute modified versions of the Photoshop binary. Most open weight model licenses allow you to make and distribute your own finetunes, etc.
I believe that comparing LLMs with traditional deterministic software is fundamentally misleading. It is extremely difficult to truly interpret what LLMs do internally, and as of now, nobody fully understands it. Even if you trained the LLM yourself, there is no source code you can simply read and learn from.
Sure, having information about how these models were trained is helpful for reproducibility, but it is basically impossible for anyone without substantial capital and access to the same (likely copyrighted) data to reproduce the model. For normal users, owning the model weights essentially means owning 100% of the model, you can inspect and study the weights in much the same way as the lab that produced the model can, you can modify the weights, and you can use and distribute them if the license allows you to
Given an open weights model trained to sometimes bite kids, we can’t train it to not bite kids, even though billions of dollars of research have been thrown at this open problem.
Given an open weights model trained to never bite kids, you can get it to bite kids with 10 prompts and a linear projection, the known simple algorithm doesn’t even need a backwards pass.
yay asymmetry!
Any pointers to more info about that? Sounds interesting.
It is useful to indicate you can run the weights on your own hardware. That’s categorically different from most other commercial offerings. It’s as if your adobe example ignores the reality that would exist had photoshop been invented in 2019: cloud only.
This analogy is terrible and seems to be extremely misinformed about how rescues evaluate dogs before they are put up for adoption
I am extremely well aware of how rescues evaluate dogs. And I'm also fully aware that they do not know the full history of the dog. They go through a limited set of testing and interrogation to evaluate the safety of the dog. That's it.