Muse Glimmer

Muse Glimmer is Meta's 30-billion-parameter open-weight model for always-on local AI agents. Released on August 10, 2026 by Meta Superintelligence Labs, it is designed to run on a high-end Mac or a single consumer GPU instead of requiring a cloud API. It accepts text and images, produces text, calls tools, writes and debugs code, follows multi-step plans, and is trained to recover when a tool action fails.

Glimmer is distilled from Meta's much larger proprietary Muse Spark model. That means Meta trained the smaller model to reproduce useful parts of Spark's behavior, then continued training it on long-context, coding, reasoning, and agent tasks. The result is not a miniature copy with identical capability; it is a local model optimized for a narrower balance of quality, memory use, and speed.

Architecture and technical specifications

Muse Glimmer is a dense causal transformer with approximately 29.6 billion parameters across 52 layers. Unlike a Mixture-of-Experts model, a dense model uses all of its language-model parameters for each token. Its attention pattern repeats three local sliding-window layers followed by one global-attention layer, reducing memory pressure while preserving periodic access to the wider sequence.

The model includes a dedicated ViT-G/14 perception encoder of roughly 1.8 billion parameters for image understanding. It accepts interleaved text and images and returns text; audio output is not supported, and video is handled as individual frames rather than through a video-native encoder. The model card lists a 131,072-plus-token context length, a 202,048-token vocabulary, grouped-query attention, SwiGLU feed-forward layers, and RoPE position encoding. Its stated knowledge cutoff is January 4, 2026.

Muse Glimmer supports low, medium, high, and xhigh reasoning strengths. It was trained on data spanning more than 100 languages, although Meta warns that it has not evaluated performance equally across all of them.

What it is built for

The model targets local personal agents that manage files, schedules, documents, and other private context; coding agents that modify and debug software; structured function calling; screenshot and chart interpretation; synthetic-data generation; and LLM-as-a-judge evaluation. It is trained for long-horizon execution, precise tool schemas, scaffold compatibility, and retrying after failed actions rather than simply stopping.

Running locally can keep prompts, documents, code, and credentials on the user's machine and removes per-token API charges. That privacy benefit depends on the surrounding application: a local model can still send data elsewhere if its tools, telemetry, or agent integrations use external services. Local weights also do not make autonomous actions inherently safe, so irreversible operations still need permission checks and application-level guardrails.

Hardware and local deployment

The full BF16 checkpoint is about 60 GB and targets systems with roughly 64 GB of GPU or unified memory. Meta also publishes approximately 4-bit quantized versions. Its K-Quant-17GB language-model weights fit under 20 GB, while the complete runtime needs a 24 GB or 32 GB memory envelope for the KV cache, image encoder, and optional speculative-decoding model. In ordinary terms, Muse Glimmer can run on one computer, but not comfortably on a typical low-memory laptop.

Meta validated quantized deployment on an NVIDIA RTX 5090 and Apple M4 Max and M5 Max systems. The release includes a small DFlash drafter model that proposes blocks of tokens for the main model to verify, improving generation speed without changing the accepted output. Meta reported 233.4 tokens per second on an RTX 5090 with this speculative-decoding setup, but speed depends heavily on context length, quantization, runtime, prompt, and hardware.

The weights are available through Hugging Face. Deployment options include llama.cpp, MLX, ExecuTorch, vLLM, SGLang, Ollama, LM Studio, and hosted third-party providers. At launch, Meta did not offer Glimmer through its own metered Model API, so there is no single official API price; hosted cost and speed depend on the provider, while self-hosting shifts the expense to hardware and electricity.

License and openness

Muse Glimmer's full-precision weights, quantized checkpoints, perception encoder, and DFlash drafter are released under the permissive Apache License 2.0. Developers can download, modify, fine-tune, and use them commercially subject to the license and Meta's usage policy.

Open weights do not mean every part of the development process is public. Meta provides architecture and methodology details, but it does not release the complete training dataset or every artifact needed to reproduce training from scratch. Artificial Analysis gave Glimmer 44 on its Openness Index, stronger than most open models but short of complete reproducibility.

Benchmark profile

Meta's official model card compares Glimmer at high reasoning with similarly sized Gemma and Qwen models.

Official model-card evaluation	Muse Glimmer 30B	Gemma 4 31B	Qwen3.6 27B
MCP Atlas (public)	75.5	54.2	62.5
DeepSearch QA	74.6	61.7	71.1
tau3-Banking	23.5	15.1	16.7
AIME 2026	94.7	89.2	94.1
GPQA Diamond	83.5	85.7	84.2
Humanity's Last Exam, text	22.0	23.6	23.1
AA-LCR	80.0	68.3	73.3
Beam128K	65.1	58.2	63.0

Meta separately reports 76% on SWE-bench Verified, 51.2% on SWE-bench Pro, and 74% on MMMU-Pro for Glimmer. Bold in the comparison table marks the highest reported score on that row, not an overall winner. These are vendor-run evaluations from Meta's model card and methodology report; prompt settings, reasoning modes, and evaluation scaffolds affect the result.

Artificial Analysis tested the released model independently and found a more mixed profile.

Independent evaluation	Muse Glimmer	Qwen3.6 27B	Gemma 4 31B	Read
Intelligence Index	35	38	30	Strong for 30B, but not the size-class leader.
GDPval-AA v2	953 Elo	1,141 Elo	Lower than Glimmer	Agentic knowledge work is a weakness.
Terminal-Bench v2.1	52%	61%	Lower than Glimmer	Qwen leads terminal-agent work.
tau3-Banking	24%	17%	—	Tool use is Glimmer's clearest independent strength.
AA-Omniscience hallucination rate	82%	49%	—	Lower is better; Glimmer is substantially worse.
MMMU-Pro	74%	—	—	Visual reasoning with the 1.8B perception encoder.

A dash means the cited independent analysis did not provide a directly comparable number in that comparison, not that the model scored zero. The independent result makes the useful verdict narrower than Meta's launch framing: Glimmer is unusually capable for a locally deployable 30B model, especially at tool use, but it is not a small frontier model across every form of agent work.

When to use it

Muse Glimmer is a strong candidate when local data control, offline availability, predictable self-hosting costs, customization, image understanding, or agentic tool use matters. It is particularly relevant for developers with 24–32 GB of GPU or unified memory who want a capable coding or personal agent without sending every prompt to a cloud model.

When not to use it

It is not the easiest choice for low-memory consumer machines, latency-sensitive services without suitable accelerators, audio-native applications, or tasks that require the strongest available cloud-model reasoning. Its weak independent hallucination and knowledge-work results mean high-stakes agents still need retrieval, verification, human approval, and narrow tool permissions.

Muse Glimmer vs. Muse Spark

Muse Glimmer is the downloadable student; Muse Spark 1.2 is the proprietary teacher. Glimmer has 29.6B parameters, a 131K-plus context window, Apache 2.0 weights, and local deployment. Spark's parameter count is undisclosed, its context window is one million tokens, and access is through Meta's paid API or Muse Code. Choose Glimmer for control and local privacy; choose Spark when higher capability and its tightly integrated cloud coding workflow matter more than self-hosting.

Bottom line

Muse Glimmer is a meaningful return to open weights for Meta: permissively licensed, technically documented, multimodal, and genuinely deployable on one high-end computer. Its practical achievement is not that it replaces the best cloud models. It is that a model small enough to keep private files and agent loops on-device can still compete with much larger systems on selected coding and tool-use tasks.