Meta open-sources Muse Glimmer 30B, an on-device agent model hitting 233 tok/s on RTX 5090
RobertTLange · x · 2026-10-25
Meta released Muse Glimmer 30B (Apache 2.0), an open-source on-device agentic model distilled from Muse Spark, designed to run locally on consumer GPUs.
Key specs
- 29.6B dense params + frozen 1.8B ViT-G/14 image encoder; text and image input
- 3:1 sliding-window attention (2048, RoPE) with NoPE global layers; GQA 32q/2kv → only 13 KiB KV/token (1.6 GiB at 128K)
- σ-gated attention, sandwich norms, logit softcap
- Training: KL on Spark's top-256 logits, then on-policy KD + RL; 3-layer draft reader with DFlash
Performance
- 17–20 GB quantized (4–5 bits/weight), 233 tok/s on an RTX 5090
- Claimed SOTA in its size class on tool use, multi-step planning and agentic completion vs Gemma4-31B and Qwen3.6-27B (MCP Atlas 75.5 vs 54.2); leads on tools and long context, trails on terminal and GUI tasks
- Dense model is 2–2.5× more compute-efficient than MoE at a fixed 32 GB budget
More from Models
- SRE Bench: GPT-6 scores 57% at binary reverse engineering, DeepSeek models under 2% — teortaxesTex · 2026-10-25
- DeepMind-backed benchmark: best AI agent solves 49/50 unpublished research-level physics problems — rohanpaul_ai · 2026-10-25
- MOSS-Image: 5B open image model hits 1.3M downloads in four days — neurocartel · 2026-10-25
- Why DeepSeek's Flash models dazzle but Pro versions disappoint: a training-signal theory — teortaxesTex · 2026-10-25
- Pro user: 62,500 Codex credits ($2,500) lock to smallest models the day your plan ends — lanqo88 · 2026-10-25
- Could OpenAI throttle Codex usage differently per user? Redditors demand transparency — Dadideology · 2026-10-25