Local & Private AI

Muse Glimmer 30B

Meta's return to open weights skips the Llama license entirely — clean Apache-2.0, on-device agents

Open ✓Model

What it is

Meta's 29.6B dense causal transformer — 27.8B of language model plus a frozen 1.8B ViT-G/14 perception encoder — distilled from its closed Muse Spark and tuned for local, always-on agent workflows: coding agents, tool calls, LLM-as-a-judge. Text-plus-image in, text out, 131K context, shipped as BF16 and two 4-bit quantizations with a DFlash speculative-decoding drafter, official ExecuTorch and GGUF builds, and day-one Ollama support.

Why it's interesting

Every prior Meta release of consequence arrived under a Llama Community License with an acceptable-use policy stapled to the grant. Muse Glimmer doesn't: the LICENSE file in the repo root is the standard, unmodified Apache-2.0 text, and the card says all artifacts share it. For the company that defined 'open-ish', shipping a genuinely OSI-approved license is the week's biggest structural news.

Use cases

  • Local always-on coding and tool-calling agents on a single consumer GPU
  • On-device multimodal assistants via the official ExecuTorch builds
  • LLM-as-a-judge evaluation pipelines that shouldn't leave your network

Who it's for

Local-agent builders, on-device developers, and Llama-license refugees

Setup

Moderate. ~60GB BF16, or the official 4-bit quants for single-GPU and Mac use; ExecuTorch PTE builds for mobile/edge, Ollama support since 0.32.7

Limitations & cautions

It's a distillation of a closed parent you can't inspect, the perception encoder is frozen and the training data unreleased, and Meta's own positioning is agentic execution rather than peak reasoning — plus a usage-policy file whose legal weight is deliberately ambiguous.

Editorial takeaway

For a decade the question was whether Meta would ever drop the Llama license. The answer arrived quietly, in an 11.4 kB file that says Apache and means it.

Related & alternatives