A week of open weights: from a Flash bump to an 816B MoE, licenses and channels matter more than parameter counts
DeepSeek, Meta, StepFun, Liquid and Microsoft AI each moved open or downloadable model assets this week. What matters for readers is how licenses, release channels and intended use differ—not treating them as one “open-source wave.”
On 17 Oct DeepSeek released V4.1-Flash-1017, a smaller update on the V4.1 line. Official notes stress stronger coding/agent benchmarks and shorter outputs; weights are MIT-licensed and a search API opened the same day. Pricing, QPS and free-period terms follow the docs. It is a faster/cheaper path for existing V4.1 capability, not a new foundation-model generation.
Meta, via a Hugging Face blog post, released Muse Spark Open weights: an ~816B multimodal MoE with a claimed 1M context. The license requires model hosts with about $20M+ annual revenue to sign a separate agreement, so open weights are not free commercial hosting for everyone. StepFun’s Step 5 Preview followed its 20 Sep announce / ~15 Oct weights-visible schedule—a preview channel, not a final GA cut.
Liquid AI shipped LFM2.5-24B-A2B open weights aimed at local agents. Microsoft AI said it finished pre-training MAI-Base-2 (~2.8T total-parameter MoE) and claimed up to ~12× training efficiency versus a DeepSeek base—vendor self-report, and finished pre-training is not the same as a public download or API. The shared theme is more model assets; licenses, local downloadability and preview status still need separate checks.
Sources
- DeepSeek-V4.1-Flash-1017: Wish-Driven Programming Made Real — DeepSeek
- Welcome Muse Spark Open: open-weight frontier multimodal mixture-of-experts model from Meta — Hugging Face
- 阶跃发布旗舰模型 Step 5 Preview：向前一步，智能效率的新一代“帕累托前沿” — 阶跃星辰
- LFM2.5-24B-A2B (Liquid AI blog) — Liquid AI
- Frontier Pre-training at Microsoft AI — Microsoft AI