alt Hacker NewsMuse Glimmer: 30B-parameter model optimized for always-on local agent workflows
Comments
The SWE bench verified score is similar to Opus from not so long ago.
Sure, you can get better performance from cloud models.
But most software, not just AI, will be faster and more reliable in the cloud. The question is do we need that additional power and cost.
If the answer is no, then just like other software, people will run AI locally.
Meta's clearly changing strategies back towards their original "frontier open source", but this time around they have a lot more competition from leading Chinese labs.
I'm all for it though, and I think Glimmer is a fantastic bet on locally-hostable models. I for one would love to self-host as much as I can.
Clear, useful, done
How many data using in Polish Language?
Qwen 3.6 27B is still such a beast!
Random question: Would you be able to run this model on a Macbook Air M5 (latest)?
LOL the mogging of GDM is hilarious.
I don't know why MSL released this, but it is very nice that they did.
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.
what's the catch?
From twitter Alexandr Wang
> 3/ muse glimmer was developed with its own architecture and recipe, optimized for its size and agentic performance requirements.
This means we're in the endgame does it not? If the architecture was NOT optimized for intelligence ...
and now the recent Meta model 'security issue' begins to make sense
thank you zuck.
what I can tell is that Meta is just starting and it is so underrated.
[flagged]
[dead]
[flagged]
[dead]
[dead]
[dead]
[flagged]
[flagged]
[flagged]
[dead]
[dead]
[dead]
Three of these landed in the same week. Mistral's Shieldstral is a 3B safety classifier that matches models 7x its size, and Google shipped Gemma Translator which runs entirely offline. Different problems, same shape. Small open weights, local, no API call.
[flagged]
"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.
Quietly the best thing in the thread
PSA: Fast RAM isn't going to be getting cheaper anytime soon. Acquiring inference hardware is a really good way to own an appreciating hard asset. Learning how to use it and cool it is a hacker's journey worth taking. My 4090 I bought in late 2022 for $1600 is selling for a cool $3,489.95 right now, and going strong under nominal use. My DRR5 has tripled in value, my nvmes almost doubled. I grabbed a 128GB M5 Max MacBook Pro when they were still available and told all my friends to buy at least one. With that and a base M4 Studio 36GB, HuggingFace rates that hardware as:
> Amazing! You have a total of 128.94 TFLOPS of computing power. 71.3% percentile on scale of "GPU Poor" to "GPU Rich"
The way I see it, these are amazing machines that the richest folks are hovering up. I think they should be in the hands of regular people as much as possible. They depend on an incredibly global, increasingly fragile supply chain. If the become impossible to produce, their value would increase tremendously. I think they will become really valuable to you to use the tokens directly, but if that isn't the case, they can be rented out or resold. Please don't just buy any hold. Let's try to get as many people that can use them for decent things that help humans. For example:
https://spectrum.ieee.org/small-language-models-ai-pharmaceu...
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.
How are you handling the tradeoff between quantization for device fit and accuracy loss on tool calling? That's where local agents typically break down in production.