Muse Glimmer
Meta Muse Glimmer: 30B open-weight model for local agents. Distilled from Muse Spark with 4-bit quantization, DFlash decoding, and tool failure recovery.
Model Overview
Capabilities, design details, and architectural traits
Muse Glimmer - Distilled Local Agent
Muse Glimmer is a 30-billion-parameter dense causal transformer distilled from Meta's closed flagship Muse Spark, purpose-built to run autonomous agentic workflows. Its training combined logit distillation from Muse Spark, mid-training on longer-context and agent-heavy data, and a post-training stage mixing supervised fine-tuning with on-policy distillation and reinforcement learning.
The model is engineered around two technical interventions that make local deployment practical. Weights are compressed to approximately 4-bit precision. Generation is accelerated by a DFlash block-diffusion drafter that proposes entire 16-token blocks in a single forward pass; the main model verifies these proposals in parallel, producing identical output quality at up to 3.1x higher throughput.
| Trait | Detail |
|---|---|
| Distilled from Muse Spark | Logit distillation from the closed flagship during pre-training, followed by on-policy distillation and RL in post-training |
| DFlash speculative decoding | Block-diffusion drafter predicts 16-token blocks; main model verifies in parallel for up to 3.1x speedup with identical output |
| Tool failure recovery | Diagnoses failed or unexpected tool-call results and retries rather than halting the workflow |
| Dedicated perception encoder | ~1.8B ViT-G/14 processes interleaved text and images including screenshots, charts, and documents |
| Controllable effort | Supports adjustable reasoning strength to balance quality against speed |
| Apache 2.0 license | Permits commercial use, modification, and redistribution without the usage restrictions attached to earlier Llama models |
Scaffold and Orchestration Compatibility
Muse Glimmer works across agentic orchestration patterns including OpenClaw and Hermes Agent, handling function calls with precise schemas across extended multi-step workflows. It is trained on data from more than 100 languages and supports a context length of 131,072+ tokens.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | Meta | Anthropic | Anthropic |
| Release Date | August 10, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | Jan 2026 | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 131K | 1M Best Context Window | 1M Best Context Window |
| Pricing (per 1M tokens) | |||
| Input Pricing | $0.33 Best Input Pricing | $4 | $2 |
| Output Pricing | $1.35 Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | textimagefile | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 17.5 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | 49.0 | - | - |
| Agentic Index | 8.5 | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Explore more from Meta
Other models by Meta
Top AI Models
Leading alternatives by intelligence score
Claude Opus 5.5
GPT-6 Astra
Gemini 4 Argon
