Muse Glimmer
Meta's 30B dense open-weight agent model — Apache 2.0, built to run always-on local agents on one consumer GPU.
About
Muse Glimmer, released August 10, 2026 by Meta Superintelligence Labs, is a 30-billion-parameter dense model engineered for local agentic work. It is distilled from the larger proprietary Muse Spark teacher and released under the Apache 2.0 license, Meta's first fully open release since the Llama era ended.
Glimmer accepts text and image inputs (via a ~1.8B-parameter ViT-G/14 perception encoder) and produces text, with 131,072+ tokens of context and training aimed at planning, function calling, failure recovery, multi-step execution, local coding, and LLM-as-a-judge evaluation. Weights ship on Hugging Face in BF16 plus two 4-bit quantized builds: K-Quant-17GB fits a 24GB card, K-Quant-Dynamic targets 32GB. A DFlash speculative-decoding drafter ships alongside.
Day-zero runtime support covers Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM, and SGLang. Knowledge cutoff is January 2026.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Provider-reported: Meta claims strong performance on agentic use cases vs leading models in its size class; 4-bit quants lose 0.2-1.0% accuracy across 15 internal benchmarks. Independent: no published independent benchmark scores at research time.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Release facts (August 10, 2026; Meta Superintelligence Labs; 30B dense; Apache 2.0; Hugging Face; distilled from Muse Spark; 131K+ context; 4-bit variants) corroborated across VentureBeat, Indian Express, Notebookcheck, and community docs. Docked for: (1) benchmark and quantization figures are Meta's own measurements with no independent assessment; (2) needs 24GB+ VRAM, so 'consumer hardware' means a 4090/5090 or M4 Max, not a mid-range card; (3) weights-only open, no training data or code.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 9.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Adoption 7: Meta's first fully open release since Llama; day-zero support across Ollama, LM Studio, llama.cpp, MLX, vLLM, SGLang
- · Maintenance 9: released 2026-08-10; weights on Hugging Face with BF16 + 2x 4-bit builds
- · Openness 10: Apache-2.0 (0.7*10) + official distribution verified at Hugging Face (0.3*10)
- · Manual 8: release facts (Aug 10 2026; 30B dense; Apache 2.0; distilled from Muse Spark; 131K+ context) corroborated across VentureBeat, Indian Express, Notebookcheck; docked for provider-only benchmark figures, 24GB+ VRAM requirement, weights-only open