Muse Glimmer model card leaks: 80.0 on AA-LCR long-context eval, beats Gemma4-31B and Qwen3.6-27B
TimDarcet · x · 2026-10-21
TimDarcet shared a capabilities scorecard for Muse Glimmer, stressing that its long-context numbers come from realistic thinking + QA evals (AA-LCR with 100k-token prompts), not synthetic needle-in-a-haystack tests.
- AA-LCR: 80.0 vs 68.3 (Gemma4-31B thinking) and 73.3 (Qwen3.6-27B thinking)
- Beam 128K: 65.1, also ahead of both
- IFBench instruction following 77.0, edging out Gemma4-31B's 76.0; SkillsBench 44.3 trails Qwen3.6-27B's 46.6
- Author claims the model agents well without compacting after every tool call
Related event: Meta's Muse-Glimmer-30B: On-Device Agent Trained on Simulated Digital Lives(9 posts)→
More from Models
- GPT-6 Astra and Claude Fable 5.1 write $10,000 research papers — experts reject all — sayashk · 2026-10-21
- GitHub Copilot's HydraFusion Orchestrates Multiple Models Per Task — WirelessLife · 2026-10-21
- OpenVLHarness turns any MLLM into a visual reasoner, +9.9 points across 22 datasets — yining_hong · 2026-10-21
- User claims Gemini analyzed his tennis match video and improved his game — zacharynado · 2026-10-21
- Opus 5.5 tops OpenRouter spending in 20 of 29 task categories — mmurph · 2026-10-21
- Meta team builds open-source 30B LLM from scratch in months, hails 'second era' — TimDarcet · 2026-10-21