Ternary-Bonsai-2 runs a 27B on your laptop at near-FP16 quality
On DevDay eve, the open-weights lane is loud: a 2-bit model for consumer GPUs, China Telecom's Apache-2.0 agent LLM, and Yandex's 80B base to fine-tune.

Copy markdown
A 27B that fits on your laptop
prism-ml's Ternary-Bonsai-2-27B squeezes Qwen3.8-27B down to ~1.76–2.1 bits per weight — a 6–7GB download that runs ~47 tok/s on an M5 Max and ~130 tok/s on an RTX 5090, while holding ~98% of FP16 quality (84.78 avg across 14 benchmarks, 96.6 math, 89.4 coding) at 262K context, Apache-2.0. Catch: it needs prism-ml's llama.cpp fork — stock llama.cpp can't load the format, though an MLX 2-bit build covers Apple Silicon.
China Telecom ships an Apache-2.0 agent model
Xing4.0-29B-A4B is a 29B mixture-of-experts that fires just 4B params per token, tuned for multi-step planning and tool calls — 256K context native (512K extended), 75.0 on SWE-bench Verified and 57.5 on Terminal-Bench 2.1. It serves on vLLM or SGLang behind an OpenAI-compatible endpoint, and the Apache-2.0 license means you can self-host it for agent work with no per-token fees.
Yandex open-sources an 80B base to build on
AliceAI-Foundation-80B-A3B-Base is a raw pretrained MoE — 80B total, 3B active, 262K context, 512 experts on top-10 routing, Apache-2.0. There's no alignment layer, so it's a clean starting point for LoRA fine-tunes, tool-use datasets, and your own instruct model via Transformers, vLLM or SGLang rather than a chat model you use as-is.