Muse Glimmer

Learning Objectives

• Understand what Muse Glimmer is and why its Apache 2.0 license is a departure for Meta
• Explain how the perception encoder and speculative-decoding drafter make an always-on local agent practical
• Evaluate when a 30 billion parameter local model is the right choice against a closed frontier API

What Is Muse Glimmer?

Muse Glimmer is Meta's open-weight model for local agent workflows, released on August 10, 2026. It is roughly 30 billion parameters — about 29.6 billion including a 1.8-billion-parameter perception encoder — with a 131,072-token context window. Meta positions it not as a frontier competitor but as the model you leave running on your own machine: watching a screen, calling tools, and handling long chains of small tasks without a network round trip.

The license is the headline. Muse Glimmer ships under a plain Apache 2.0 license — no monthly-active-user threshold, no geographic carve-out, no attribution clause, and no acceptable-use addendum layered on top. Every prior capable Meta release, including the entire Llama 4 family, carried the Llama Community License and its 700-million-user ceiling. Apache 2.0 puts Glimmer on the same footing as Gemma 4, Mistral Small 4, and Qwen — genuinely permissive, and unusual enough from Meta to be the story.

Why "Always-On Local" Is the Design Goal

Most open-weight releases are scaled-down versions of a frontier model, and they are evaluated as such. Glimmer is built backwards from a deployment constraint: it has to fit on hardware you already own and stay resident without making your machine unusable.

Two engineering choices follow from that.

Quantization to a Consumer Memory Envelope

At full precision the language model needs over 55 gigabytes. Quantized to roughly 4-bit precision, it drops to under 20 gigabytes — which is what lets it sit inside a 24- or 32-gigabyte envelope on a consumer GPU or an Apple Silicon Mac. That is the difference between a model you rent by the hour and one that runs on a single desktop card.

A Speculative-Decoding Drafter

Glimmer ships with a drafter built on Meta's DFlash technique. A small companion network proposes whole blocks of tokens at once and the main model verifies them in parallel, so output is identical to normal decoding but arrives faster. Meta reports 3.1-times faster decoding on an RTX 5090, 1.8-times on an M5 Max, and 1.5-times on an M4 Max. For an agent that runs continuously rather than answering one question at a time, decode speed is the whole experience.

A Perception Encoder for Screens

The 1.8-billion-parameter perception encoder lets Glimmer read interleaved text and images — screenshots, charts, and documents inside a conversation. This is the capability an on-screen agent actually needs, and it is why the encoder is bundled rather than offered as a separate vision model.

Benchmarks

Meta evaluated Glimmer against Google's 31 billion parameter Gemma 4 and Alibaba's 27 billion parameter Qwen 3.6 — its actual size class, not the frontier.

All figures are vendor-reported and were published alongside the release, so no independent replication exists yet. Treat the coding numbers as the meaningful ones for this model's purpose — a 76 percent SWE-bench Verified score from a model that fits on one consumer card is the claim worth checking against your own workload.

Pricing and Access

The weights are free under Apache 2.0. The only cost is hardware you already own, or per-token pricing if you route to a hosted provider instead.

Strengths

• A genuinely permissive license: Apache 2.0 with no user threshold, no territory exclusion, and no acceptable-use addendum — you can build a competing commercial product on it without asking Meta
• Fits real consumer hardware: Under 20 gigabytes at roughly 4-bit precision, inside a 24-gigabyte card
• Fast decode where it matters: Up to 3.1-times faster generation through the DFlash drafter, with identical output
• Reads screens natively: The bundled perception encoder handles interleaved text and images without a separate vision pipeline
• Serious tool calling: Handles precise function schemas across extended workflows, with scaffold compatibility for OpenClaw and similar orchestration patterns
• Wide runtime support at launch: Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM, and SGLang, plus three hosted providers

Limitations & Considerations

• Not a frontier model, by design: At 30 billion parameters it will not match a closed flagship from OpenAI or Anthropic on the hardest reasoning and long-horizon coding work — the trade is capability for locality
• Benchmarks are vendor-reported: Published by Meta at launch with no third-party replication yet
• The 4-bit figure is the deployable one: Full precision needs over 55 gigabytes, so the consumer-hardware story depends on quantization, which costs some quality on the hardest tasks
• Drafter speedups are hardware-specific: The 3.1-times figure is an RTX 5090 number; Apple Silicon sees 1.5 to 1.8-times
• Meta's open-weight track record is uneven: Llama 4 Behemoth was deprioritized and the community license has shifted before, so treat the Apache 2.0 posture as current policy rather than a permanent guarantee

Best Use Cases

When to choose alternatives:

• Maximum reasoning capability → a closed frontier flagship from OpenAI or Anthropic
• Smallest possible footprint → Phi-4 Mini or Gemma 4 E2B for phones and embedded devices
• Meta's most capable model overall → Muse Spark, which is proprietary and API-only
• A longer-established open ecosystem → Qwen or Gemma, which have more community fine-tunes today

Getting Started

1. Confirm your hardware: a 24-gigabyte GPU or a 32-gigabyte Apple Silicon Mac is the target envelope
2. Install Ollama or LM Studio and pull a 4-bit quantized Glimmer build
3. Verify your runtime supports the DFlash drafter — without it you lose the decode speedup that makes always-on use comfortable
4. Test tool calling against your own function schemas before wiring it into an agent loop
5. For screen-reading workflows, confirm your runtime exposes the perception encoder's image path
6. Benchmark against your real tasks rather than the published scores, which are vendor-reported
7. If local hardware is the blocker, try it hosted on OpenRouter first, then move local once the workflow is proven

Key Takeaways

• Muse Glimmer is Meta's 30 billion parameter open-weight model for local agents, released August 10, 2026 under a plain Apache 2.0 license with no user threshold or geographic carve-out
• That license is a break from Meta precedent — every prior capable release, Llama 4 included, used the Llama Community License and its 700-million-user ceiling
• Quantized to roughly 4-bit it runs under 20 gigabytes, fitting a 24-gigabyte consumer GPU, and the DFlash drafter delivers up to 3.1-times faster decoding with identical output
• A bundled 1.8-billion-parameter perception encoder reads interleaved text and images, which is what an on-screen agent needs
• Vendor-reported benchmarks put it at 76 percent on SWE-bench Verified against same-size peers Gemma 4 and Qwen 3.6 — strong for the class, well short of a closed frontier flagship