Meta opens Muse Glimmer 30B for local agents (Apache 2.0)

Meta Superintelligence Labs open-sourced **Muse Glimmer** on 10 August 2026 under Apache 2.0. About 29.6B parameters with a dedicated perception encoder, built for always-on local agents on one consumer GPU. Weights: meta-models/Muse-Glimmer-30B. Primaries: the Meta research blog, the Hub card, and developer docs.

From @AIatMeta on X (10 August 2026 UTC):

"Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows. Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category, and is designed to run entirely on…"

The blog trains Glimmer from Muse Spark via logit distillation, then agent-heavy mid- and post-training (SFT, on-policy distillation, RL). The card lists a dense causal transformer plus a ~1.8B ViT-style perception encoder inside the ~29.6B total, text-plus-image in, text out, 131,072+ context. Meta says ~4-bit quantization puts the language model under 20 GB so KV cache, vision encoder, and a DFlash speculative-decoding drafter fit a 24-32 GB envelope. llama.cpp, MLX, ExecuTorch, and partner paths are listed for the coming days.

Key facts

• Live Hub: meta-models/Muse-Glimmer-30B (ungated on public API; Apache 2.0 plus separate USAGE_POLICY)
• Scale: ~29.6B total including ~1.8B encoder; BF16 safetensors about 29.8B parameters across two shards
• Local pitch (Meta-reported): single consumer GPU; quantized LM under 20 GB with DFlash in a 24-32 GB envelope
• Training targets: tool calling, multi-step plans, failure recovery, controllable effort, 100+ languages
• Eval names: DeepSearch QA, MCP-Atlas, tau3-Bench family, SWE-Bench; peers Gemma4-31B and Qwen3.6-27B
• Family: open local sibling distilled from Muse Spark; Spark stays the larger Muse API track on this ship day

How to read this

Hub files and the Apache license are the hard download fact. Benchmark cells, "strong for size class," under-20 GB claims, and scaffold names (including OpenClaw and Hermes Agent) are Meta-reported. USAGE_POLICY is separate from the Apache grant.

Independence

Frontier Flux is not affiliated with Meta, Meta Superintelligence Labs, or Hugging Face. This note follows the public blog, Hub card, license, usage policy, and @AIatMeta announcement.