Meta open-sources Muse Glimmer 30B, an on-device agent model that beats Gemma4-31B on MCPAtlas
dair_ai · x · 2026-10-20
Meta Superintelligence Labs released Muse Glimmer, an open-source 30B model distilled from its frontier model Muse Spark, purpose-built for autonomous agentic tasks running entirely on consumer hardware under Apache 2.0.
Key facts
- With 4-bit quantization and a speculative decoding head, it fits in 32 GB of VRAM and runs on a single RTX 5090
- Accepts text and image inputs, trained on 100+ languages
- MCPAtlas: 75.5 vs 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B
- SWE-Bench Pro: 51.2 vs 36.9 and 50.2
- Qwen3.6-27B still leads on OSWorld-Verified and TerminalBench 2.1 — the gains are concentrated in tool use and long-horizon tasks
The paper argues targeted distillation from a frontier teacher yields capable small models at lower inference cost, and that a dense architecture beats MoE under the 32 GB memory constraint. Worth bookmarking for anyone building locally-running agents.
More from coding & agent
- How engineers rescued a vibe-coded food-safety app serving KFC and Taco Bell kitchens — alex_verem · 2026-10-20
- FrankenTerm going native on iPhone and iPad: terminal mux poised to manage dev machines on the go — doodlestein · 2026-10-20
- From localhost to production: deploy an AI agent in 3 commands with 11-second responses — Arindam_1729 · 2026-10-20
- Databricks Genie Agents turn natural language into SQL analytics — adnan_hashmi · 2026-10-20
- "The End of Software Engineering" paper plus a guide to cut Claude bills from $765 to $3 — Roger_M_Taylor · 2026-10-20
- Ethiack Zero launches: a public live record of emerging 0-days with API and MCP — rez0__ · 2026-10-20