While Mistral Large 4 is a good news for Europe and AI, it is not a good fit for DwarfStar because of the very large number of active parameters, not just because of memory: would be too slow anyway. But anyway congrats to @MistralAI and waiting for a smaller MoE :D
Oct 21, 2026 · 8:57 PM UTC
11
13
369
19,010
@antirez @MistralAI You don’t have to wait! We released a 78B-A3.5B MoE ~3 weeks ago, fully European and Apache 2.0. Dovresti provarlo :)
huggingface.co/Aleph-Alpha/Ko…
huggingface.co/Aleph-Alpha/Ko…
2
0
14
800
@alessio__serra @MistralAI ho visto il rilascio ne sono felice, ma per il progetto ha senso solo supportare modelli nella frontiera di Pareto di intelligenza su dimensione.
1
0
8
681
@antirez @MistralAI Good point!
What about 2.5 billion active parameters out of 39.8 billion total parameters of the nice Flint Recursive’s mixture-of-experts?
What about 2.5 billion active parameters out of 39.8 billion total parameters of the nice Flint Recursive’s mixture-of-experts?
0
0
0
79
@antirez @MistralAI DeepSeek v4.1 Flash will do in the meantime 😁
0
0
0
80
@antirez @MistralAI What is the ideal current DwarfStar MoE total params and active params based on what DwarfStar users are using as hardware right now?
0
0
0
646
@antirez @MistralAI 49b active vs the 13-18b ds4 is built around, yeah that was never fitting. weights dont even ship till end of october anyway
0
0
0
594
@antirez @MistralAI ~49B active per token is a lot, a smaller Mistral MoE would be the one I'd actually try
0
0
0
223
@antirez @MistralAI The cat is too fat 😂 Looking forward to ds support for future small models 👀
0
0
1
651