5 ms·
Okay but Muse Glimmer 30B is one of the best small open weight models today, and IMO the best from a US lab (only real comparison is Gemma4 dense right now).
by aftbit 14d ago
Okay but Muse Glimmer 30B is one of the best small open weight models today, and IMO the best from a US lab (only real comparison is Gemma4 dense right now).
- Bluestein 14d agoI am finding Poolside's a decent model.-
- hadlock 14d agoBy their own benchmarks it is about 10% lower scoring than Qwen 3.6 35b-a3b, but I've added it to my list. Always looking for MoE to compare to it so we can squeeze more out of our local LLM system.
- Bluestein 14d agoI found it has some "tail" errors, wherein it would make up important details (ie. "happypath.exp" vs "happypaws.exp" and then claim your "DNS is having issues" - where the second domain does not exist), things like that.- ... but correctly supervised it does get some things done.-
- aftbit 13d agoI've found that's generally true of smaller / weaker models. They're quite capable, but you need to distrust them a lot and give them very detailed instructions. Even the free Gemini in Google Search is like this - it lies a lot, clips off important info, and generally goes off the rails if you do too many turns, but it's still very useful if you keep all that in mind.
- Bluestein 13d agoSpot on. Guardrails is the game.-
- tyre 14d agoTotally fine with open weight, since other people can provide it and Meta isn’t making money. I’d use an AWS-hosted version.
- HDBaseT 14d agoZuck said Muse Spark 1.2 should be getting open weights "soon" on a tweet from a few weeks ago. The problem inference providers will not be able to get anywhere near the contributor pricing.
- cheema33 13d ago> The problem inference providers will not be able to get anywhere near the contributor pricing. Maybe if they started collecting data..