Muse Glimmer

// Overview

Muse Glimmer is Meta Superintelligence Labs' open-weight agentic model, released on August 10, 2026 under the Apache 2.0 license. It is a ~29.6B-parameter dense causal transformer with a dedicated ~1.8B-parameter ViT-G/14 perception encoder, distilled from Muse Spark using logit distillation during pre-training, and aimed at always-on local agent workflows rather than hosted frontier serving.

Meta trained and evaluated it on the capabilities agents actually need: end-to-end task completion, reliable schema-based tool calling, multi-step reasoning over long horizons, recovery after a failed tool call, and interleaved text-and-image input. The model takes text and images and returns text (audio is not supported), has a 131,072-token context length, a January 4, 2026 knowledge cutoff, and training data spanning more than 100 languages. Reasoning strength is selectable in the system prompt at low / medium / high / xhigh.

Local deployment is the design goal. Meta ships 4-bit quantized weights that shrink the language model to under 20 GB — Meta reports 0.2% average degradation for its K-Quant-Dynamic build (32 GB VRAM target) and 1.0% for K-Quant-17GB (24 GB target) across 15 benchmarks — plus a DFlash block-diffusion drafter for speculative decoding. With that drafter, Meta measures 233.4 tok/s on an RTX 5090 (3.1× over its 74.9 tok/s baseline), 50.2 tok/s on an Apple M5 Max (1.8×), and 37.8 tok/s on an M4 Max (1.5×). Full-precision BF16 weights, both quantized variants, the drafter head and the frozen perception encoder are all released under Apache 2.0.

// Benchmarks

This model's scores

1. MCP Atlas (Public) — general agentic75.5
2. DeepSearch QA74.6
3. SWE-Bench Verified (agentic coding)76
4. SWE-Bench Pro51.2
5. AIME 202694.7
6. GPQA Diamond (AA)83.5

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

The weights are free under Apache 2.0 and can be run locally at no per-token cost; these are OpenRouter's hosted rates for meta/muse-glimmer-30b.

Pricing source ↗

// Strengths

• Apache-2.0 open weights — full BF16 checkpoints plus two 4-bit quantized builds, the DFlash drafter, and the vision encoder
• Runs locally on a single consumer GPU: under 20 GB quantized, validated on 24 GB and 32 GB VRAM envelopes and on Apple M4/M5 Max
• Leads its size class on Meta's agentic evaluations — MCP Atlas 75.5, DeepSearch QA 74.6, Gaia2 43.3 and τ3-Banking 23.5 against Gemma4-31B and Qwen3.6-27B
• Speculative decoding via the bundled DFlash drafter gives a measured 3.1× decode speedup on an RTX 5090
• Multimodal input through a dedicated perception encoder — screenshots, charts and documents alongside conversation
• Selectable reasoning strength (low / medium / high / xhigh) set from the system prompt

// Best for

• Reach for it when an agent must run entirely on-device — no cloud round-trip, no network dependency — with tool calling and failure recovery intact.
• Reach for it for local coding agents and SWE-Bench-style workflows on a 24–32 GB consumer GPU or an Apple silicon laptop.
• Reach for it as an LLM-as-a-judge evaluator or a synthetic-data generator where per-token API cost would otherwise dominate.
• Use Muse Spark 1.2 instead when the job needs frontier-level quality, a 1M-token context, or Meta's hosted API rather than local weights.

// How to access