Official OsaurusAI affine-quantized build of poolside/Laguna-M.1 (Apache-2.0) — a 225B / 23B-active agentic-coding MoE. Quantized by Osaurus to the JANG_2L profile (~81 GB), runs on a single 128 GB Apple-Silicon machine via Osaurus / vMLX.

Profile (JANG_2L): 2-bit gate/up, 3-bit down, 6-bit shared/dense, 8-bit attn/embed-head — quality 2-bit (affine, group-size 64; per-module bits in config.json).

Requires a Laguna-aware runtime (Osaurus / vMLX / jang_tools.laguna) — Laguna is a custom architecture (per-element attention gating, YaRN RoPE, sigmoid+bias MoE routing, shared expert) not in stock loaders. All-full attention → standard KV cache. For clean chat, stop on </assistant> / 〈|EOS|〉 and use repetition_penalty≈1.15.