Meta open-sources Muse Glimmer, a 30B agent model that runs local
Apache 2.0, one consumer GPU, 131K context — and it beats Qwen and Gemma at agentic tool-calling. Plus: MiniMax-H3 video+audio now runs on a Mac.
Copy markdown
One consumer GPU, Apache 2.0, no strings
Muse Glimmer is a dense 30B (with a ~1.8B vision encoder) distilled from Muse Spark, quantized to under 20GB so it runs on a single consumer GPU — a MacBook M4/M5-Max or an RTX 5090. It ships under Apache 2.0, far more permissive than Llama's community license, so you can build on it commercially without the old usage caps.
Built to drive agents, and the scores show it
It's tuned for tool use: 75.5 on MCP-Atlas versus Qwen3.6-27B's 62.5 and Gemma4-31B's 54.2, plus 74.6 on DeepSearch QA and 51.2 on SWE-Bench Pro. It trails Qwen on raw coding (SWE-Bench Verified, TerminalBench 2.1), so reach for it as an orchestrator, not a heads-down code model.
Fast enough to sit in the loop
On an RTX 5090 it runs 74.9 tok/s greedy, ~233 with its DFlash 16-token speculative decoding; an M5 Max lands 26–50 tok/s. It takes interleaved text and images, a 131K context window, and data from 100+ languages.
Build this weekend: a private always-on agent
It's on Ollama, LM Studio, vLLM, Together and Fireworks day one. A 30B that fits in 20GB and calls tools well is enough to run a local research-or-files agent that never leaves your machine — no per-token bill, no data egress.
Elsewhere: MiniMax-H3 video+audio now runs on a Mac
A community MLX port (alongside antirez's h3.c engine) brings MiniMax-H3's 33B joint video+audio diffusion to Apple Silicon natively — generate clips with synced audio locally, no cloud. It hit the Hacker News front page at 333 points.