NaiveAI open-sources Naive-N0.5-Flash: 309B, MIT, 1M context
No full-attention layers keep long context cheap to serve, and the hosted API runs $0.10/$0.40 per M. Plus: DSPy finally ports to Elixir's BEAM.

Copy markdown
A 309B you can actually own — MIT, callable for cents
NaiveAI dropped Naive-N0.5-Flash open-weights under MIT: 309B total, 15.5B active, native 1M context, built from Xiaomi's MiMo-V2.5 base. Rent it at $0.10/$0.40 per M tokens, or pull the ~315GB FP8 weights and self-host — commercial use included.
The trick: not one full-attention layer
48 layers — 39 sliding-window (128-token) plus 9 DeepSeek Sparse Attention (top-2,048) in a 5:1 mix, with zero full attention. That's what keeps the 1M window affordable to serve; NaiveRT pushes up to 2,000 tok/s in Ultrafast mode, and an FP8 draft model ships for speculative decoding.
Weekend build: DSPy lands on the BEAM
Imp v0.5 ports DSPy to Elixir — signatures, GEPA-style optimizers, ReAct/MCP agent loops and program-of-thought, all on OTP's actor model. Needs Elixir 1.19+; a clean path to self-improving LLM programs with real concurrency and supervision.