Laya-MLX
⚠️ Why not a model
Native MLX port of ConvAI's Laya: converted FP16 checkpoints and an inference engine for Apple Silicon. No decision training of its own.
Technical specs
- Base LLMs
- Decision types
- choice, score, noul
- Features
- local, apple-silicon, mlx, pypi, http-serve, mcp, router, calibration, langchain, langgraph, llamaindex, crewai, agno
- License
- Apache-2.0
Links
publicado (OK do Fabricio 20/10 BRT).
Laya-MLX (Apache-2.0, pip install laya-mlx) runs Laya natively on Apple Silicon with MLX: no PyTorch, no Transformers runtime, zero output tokens. Pre-converted FP16 checkpoints are on Hugging Face (aac6fef/laya-mlx, laya-multilingual-mlx, laya-typed-decisions-mlx). Release 0.5.0 (pushed 20 Oct) migrates to upstream Laya v0.4.12 and adds the router, batch and long-input prediction, hooks, calibration, evaluation, HTTP and MCP serving and framework extras (LangChain, LangGraph, LlamaIndex, CrewAI, Agno).
Author numbers (UnverifiedClaim), M3 Max: 13.4 ms median end-to-end for a short English decision and 7.4 ms with the multilingual checkpoint; all three checkpoints matched upstream’s selected answer on 63 of 63 validation questions in FP32 and FP16. The repo publishes no accuracy comparison with Jev. A post on X on 19 Oct (Kim Maresana) asked exactly that; we found no answer.
Why it is here, not in the catalog. It converts and serves ConvAI’s weights; quality is Laya’s. Same class as Ollaya, which also runs Laya on MLX.