Open Weights and Cheap Models: The Quiet Story of October 2026
While everyone watched the flagship chat models, the runnable ones got better and cheaper.
Mistral Large 4, previewed October 6, is a trillion-parameter multimodal mixture-of-experts model with open weights promised by month's end. At 38 on the Intelligence Index it sits near GPT-6 Luna and DeepSeek V4.1 Flash at max effort, and its Cyber Index score of 50 is notable for a non-US model.
DeepSeek V4.1 Flash compressed the China-US gap to about 3 percent, and at extremely low cost it makes 'just use a frontier model' a harder default to defend for high-volume, low-stakes calls.
The decision-model wave kept going: tiny typed-answer models from OpenAI, Cloudflare, Perplexity, Microsoft, and Liquid AI return a choice or score in one pass for fractions of a cent, and AWS open-sourced Strands Decider 2B so you can run it yourself.
Google open-sourced EmbeddingGemma 2, a roughly 740M-parameter multimodal embedding model under Apache 2.0, which puts on-device retrieval within reach for smaller teams.
The takeaway: 'free to run' stopped meaning 'worse.' For most production calls, the question is shifting from which giant model to which small one you can host.
Updated October 28, 2026: Mistral Large 4's open weights were promised by month's end - check the latest roundup for whether they landed. Everything else below is live.
Frequently asked questions:
Q: Which open-weight models shipped in October 2026?
A: Mistral Large 4 opened in preview on October 6 with weights promised by month's end, DeepSeek V4.1 Flash stayed downloadable, Google open-sourced EmbeddingGemma 2 under Apache 2.0, and AWS open-sourced Strands Decider 2B.
Q: How close is DeepSeek to US frontier models?
A: V4.1 Flash narrowed the China-US gap to about 3 percent while costing far less, which makes 'just use a frontier model' harder to defend for high-volume, low-stakes calls.
Q: Are cheap decision models production-ready?
A: For routing, classification and typed answers, yes. OpenAI, Cloudflare, Perplexity, Microsoft and Liquid AI all ship them, returning choices in one pass for fractions of a cent.