NaiveAI open-sources a 309B MoE with 1M context, MIT-licensed
Built from Xiaomi's MiMo-V2.5 and served at $0.10/$0.40 per Mtok; plus MiniMax's stealth coding model, a 42x llama.cpp drafting win, and DSPy for Elixir.

Copy markdown
A frontier open model with no full-attention layers
Naive-N0.5-Flash is a 309B-parameter MoE (15.5B active) fine-tuned from Xiaomi's MiMo-V2.5, using 39 sliding-window layers plus 9 DeepSeek-sparse layers to reach a native 1M-token context with zero full-attention layers. MIT-licensed weights (and an FP8 build) are on Hugging Face; the hosted API runs $0.10/$0.40 per million input/output tokens.
MiniMax slips a fast coding model into its agent
MiniMax M3.1-Flash-Preview quietly landed inside MiniMax Code — a lighter, faster sibling of M3 tuned for everyday coding, with five reasoning tiers (low through max). Specs and pricing are still unpublished, and MiniMax is reportedly running 2x daily check-in credits from Sep 28 to Oct 7 to pull builders in.
A ~42x speedup for llama.cpp prompt-lookup drafting
New optimization passes reportedly cut prompt-lookup drafting from ~165µs to ~1.2µs per token (~42x), and static cache loading from 3.76s to 0.23s. The gain is workload-specific — biggest on repetitive, long-context prompts — but it's free speed for anyone running local inference.
DSPy comes to Elixir: Imp v0.5 ships on Hex
Imp 0.5.0 ports DSPy to the BEAM: typed signatures, GEPA-style optimizers, and agents that run as OTP processes with ReAct and MCP support. If you build on Elixir, you can now write self-improving LLM programs with the actor model's concurrency baked in. Needs Elixir 1.19+ and any ReqLLM-supported provider.