Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

research.meta.ai

Post reply on HN
Print view View original

121–130 of 682 posts

Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

#121

Post by anon373839 »

OtherShrezzing wrote: post #77
cynicalsecurity wrote: post #69

Earlier quoted context omitted.

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. O…

> 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.

Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

#122

Post by Grombobulous »

larodi wrote: post #109
avaer wrote: post #101

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 Met…

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…

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.

Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

#123

Post by cube00 »

cmiles8 wrote: post #97

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 [1] and loose [2] with limits, billing [3] and adding undisclosed changes that burn your tokens on autopilot [4], it can't happen soon enough.

[1]: Limits may change without notice, including due to capacity constraints. - https://support.google.com/gemini/answer/16275805?sjid=14713....

[2]: "standard limits" are never defined - https://support.google.com/gemini/answer/16275805?sjid=14713...

[3]: https://tobyonfitnesstech.com/blog/anthropic-refund-scam/

[4]: https://news.ycombinator.com/item?id=48947776

Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

#124

Post by mapontosevenths »

Manfrednotfunny wrote: post #81
mapontosevenths wrote: post #75

Earlier quoted context omitted.

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,…

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.

Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

#125

Post by mapontosevenths »

dist-epoch wrote: post #90
mapontosevenths wrote: post #75

Earlier quoted context omitted.

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,…

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.

Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

#126

Post by Forgeties79 »

echelon wrote: post #98
Forgeties79 wrote: post #94

Earlier quoted context omitted.

>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.

I don’t start counting since we WaveNet or BERT so there you go!

Even if I did, we’re talking barely a decade

Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

#127

Post by Gigachad »

dbbk wrote: post #58
sajithdilshan wrote: post #10

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.

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.

The models people are spending thousands on require more on the range of 600-800gb memory.

128gb hardly runs deepseek v4 flash which is almost free via api pricing.

Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

#128

Post by moron4hire »

"Meta Muse" immediately made me think of Metamucil.

Product teams really need to hire at least one or two people with a 12-year-old's sense is humor. They need to winnow all the potential stupid jokes out of their product namings.

Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

#129

Post by Gigachad »

IMTDb wrote: post #31
zoobab wrote: post #23

Earlier quoted context omitted.

"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.

All those random api providers are absolutely scooping up your data though. And the hardware to run it locally is absurdly expensive.

Re: Muse Glimmer: 30B-parameter model optimized for always-on local agent workflows

#130

Post by swiftcoder »

monster_truck wrote: post #119
avaer wrote: post #101

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 Met…

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 b…

> FB at one point had a very badass engineering culture

Perpetually kneecapped by one of the worst management cultures I've ever seen

Post reply on HN