Muse Glimmer

Muse Glimmer is an open-weight text-and-image model developed by Meta AI for local agent and coding workloads. Meta released the model on August 10, 2026 under the identifier meta-models/Muse-Glimmer-30B. Its public BF16 checkpoint contains 29,776,626,688 parameters according to the repository's safetensors metadata, while Meta rounds the size to about 29.6 billion parameters in the model card.^[1][2][4]

Meta describes Muse Glimmer as a dense large language model with a separate perception encoder, distilled from Muse Spark.^[2] It accepts text and images and produces text. Meta positions it for AI agents that run on user-controlled hardware, but the checkpoint is not an autonomous application by itself. A working agent still needs a runtime, tool definitions, permissions, and an orchestration scaffold.

The release includes ungated BF16 weights, two official four-bit GGUF variants, a quantized multimodal projector, and a DFlash drafting model. These are released artifacts, unlike several runtime and hosting integrations that Meta said would become available after launch.^[1][5] The Associated Press independently reported the release and its personal-computer positioning, but did not independently test Meta's performance claims.^[8]

Release and model identity

The model card names Meta Superintelligence Labs as the developer and August 2026 as the release month. Meta's announcement gives the public release date as August 10, 2026. The main Hugging Face repository is public and ungated, identifies the pipeline as image-to-text conversation, and labels the checkpoint Apache-2.0.^[1][2][4]

The "30B" name is a size class rather than an exact count. Repository metadata reports 29,776,626,688 BF16 parameters. Meta's rounded total includes the perception encoder. The public materials do not provide a separate active-parameter figure, so the total should not be restated as an independently disclosed active count.^[2][4]

Meta marketed Glimmer as an "always-on local" model for commercial and research uses including local assistants, coding agents, schema-based function calls, multimodal document or screen interpretation, synthetic-data generation, and model evaluation.^[2][7] Those are intended uses and training targets. They do not guarantee successful tool use in every scaffold or make real-world actions safe without deployment controls.

Architecture and context

Muse Glimmer combines a dense causal transformer with a dedicated vision transformer. The text stack has 52 layers with a repeating pattern of three local-attention layers followed by one global-attention layer. Local attention uses a 2,048-token window. Its 32 query heads and two key-value heads implement grouped-query attention at a 16:1 ratio.^[2][3]

Scroll sideways for more →

The model card writes the context length as "131,072+," but the released configuration fixes max_position_embeddings at 131,072.^[2][3] That configured value is the supported figure established by the artifact. Longer contexts demonstrated with community RoPE overrides are experimental extensions, not part of the official configuration.

The perception component is based on Meta's Perception Encoder family, described in a 2025 research preprint by Daniel Bolya and colleagues.^[10] Muse Glimmer can receive interleaved text and images, including screenshots, charts, and documents. Its output remains text. Audio input and output are unsupported. Although its configuration contains a video token, the model card says video was not explicitly optimized and is processed as individual frames.^[2][3] It is therefore a multimodal understanding model, not a video generator.

Agent behavior and controls

Meta trained and evaluated the model for multi-step task completion, function calling, failure recovery, and compatibility with agent scaffolds. The card names OpenClaw and Hermes Agent as examples and defines four requested reasoning-strength levels: low, medium, high, and xhigh.^[2] These controls are prompt-level instructions supplied to the model. They do not establish a fixed compute budget or a universal quality increase.

A local agent deployment can keep prompts and tool results on the user's machine, but local execution does not automatically guarantee privacy. A scaffold may still call remote search, storage, or execution services, and its tool permissions determine what the model can access. Meta recommends system-level guardrails and human confirmation before irreversible actions.^[2]

At launch, Meta linked the downloadable model and developer documentation. The announcement also said support would arrive "in the coming days" through Ollama, LM Studio, Unsloth, llama.cpp, ExecuTorch, MLX, vLLM, SGLang, Together, Fireworks, and OpenRouter.^[1] That list is an integration roadmap from the publisher, not evidence that every named runtime or hosted service was operational on August 10.

Released artifacts and local deployment

The primary repository holds two BF16 safetensors shards totaling about 59.6 GB in decimal storage units. The official GGUF repository holds two target-model quantizations, plus separate files for the perception projector and DFlash drafter.^[4][5]

Scroll sideways for more →

The 24 GB and 32 GB figures are Meta's deployment targets, not universal minimum requirements. Available context, key-value cache precision, runtime buffers, batching, offload, operating-system use, and the selected projector or drafter all change actual memory demand. A claim that the model runs on "18 GB" usually counts only one target-model file and leaves out other components and runtime memory.^[2][5]

Meta reported average accuracy degradation of 1.0 percent for KQuant-17GB and 0.2 percent for KQuant-Dynamic against full precision. The model card says these are averages across 15 common benchmarks but does not identify the complete set or publish each benchmark's result.^[2] The figures therefore support a vendor-reported aggregate comparison, not a guarantee that every task changes by those percentages.

DFlash speculative decoding

The optional drafter implements DFlash, a speculative decoding method that proposes a block of tokens with a lightweight block-diffusion model and lets the target model verify them in parallel. Jian Chen, Yesheng Liang, and Zhijian Liu introduced DFlash in a paper accepted at ICML 2026.^[9] Muse Glimmer's released drafter has five layers, proposes blocks of 16 tokens, and receives hidden features from five layers of the target model.^[2]

Meta measured the KQuant-17GB target and quantized drafter at batch size one with greedy decoding. It used llama.cpp on an Nvidia RTX 5090 and ExecuTorch on Apple systems.^[2]

Scroll sideways for more →

These measurements are specific to Meta's prompt set, quantization, backends, batch size, and decoding mode. They do not predict speed under sampling, long prompts, concurrent requests, a different cache configuration, or other hardware. DFlash preserves the target decoder's selected output only when speculative verification accepts or corrects every proposal as designed; it does not improve the target model's underlying accuracy.^[9]

Evaluation and score provenance

Meta compared Muse Glimmer with Gemma4-31B and Qwen3.6-27B in thinking modes. The publisher's evaluation report states that Glimmer used temperature 1.0, top-p 0.95, and top-k 64. For comparison models, Meta selected the more favorable value between a publisher's self-reported result and Meta's internal reproduction. It used Artificial Analysis values when that source had results for all compared models.^[6] The table therefore mixes score provenance and is not a wholly independent, single-harness evaluation.

Selected results show both strengths and losses in Meta's own table.^[2][6]

Scroll sideways for more →

For MCP Atlas, Meta ran 500 public tasks four times and used Gemini 2.5 Pro as a judge. Its SWE-bench setup used 500 Verified tasks and 731 Pro tasks with four runs and a bash-and-file tool environment; the Qwen Verified value was self-reported while the other displayed values were internal. OSWorld excluded Google Drive tasks and exposed different action interfaces to different models. The methodology report also warns that third-party agent frameworks and system prompts may not have been tuned equally for the comparison models.^[6]

The results do not justify a categorical claim that Glimmer is best in its size class. Meta's table gives Qwen higher scores on GDPVal-AA v2, SkillsBench, OSWorld, SWE-Bench Verified, TerminalBench, ScreenSpot Pro, OmniDocBench, and MMMU Pro. Gemma leads the displayed GPQA Diamond and HLE Text results.^[2] AP's independent reporting confirms that the model was released, not that the benchmark ranking was replicated.^[8]

Training disclosure, license, and safety

Meta discloses training-data sources only at a broad level: publicly available multimodal material, third-party-provided data, and information from Meta products and services, with curation and enrichment by external vendors and Meta personnel. The card says training covered more than 100 languages and that the model was distilled from Muse Spark.^[2] It does not identify datasets, mixture proportions, total training tokens, training compute, or a complete distillation and post-training recipe.

The checkpoint, GGUF weights, drafter, and perception artifact are distributed under Apache License 2.0.^[2][5] This permits broad commercial modification and redistribution subject to the license's notice, attribution, patent, and trademark provisions. Meta separately publishes a Muse Glimmer Usage Policy that prohibits categories of harmful, illegal, deceptive, privacy-invasive, and high-risk use.^[11] The usage policy is a separate publisher policy and should not be confused with the repository's Apache license label.

Meta says the model received safety supervised fine-tuning, safety reinforcement learning, and synthetic training for "appropriate information flows." It assessed content safety, agentic risk, privacy, and preparedness.^[2] Meta classified chemical and biological risk as Moderate or lower based on Glimmer evaluations. It also assigned Cyber and Loss of Control designations of Moderate or lower, but explicitly inferred those two from Glimmer being broadly weaker than Muse Spark rather than from a separate direct Glimmer assessment in those domains.^[2]

The publisher's own safety table shows unresolved vulnerabilities. On Siren AgentDojo, Glimmer had a 28.4 percent attack success rate, compared with 25.6 percent for Gemma4 and 40.3 percent for Qwen3.6. On CI Memories, Glimmer's violation rate was 26.4 percent, compared with 12.1 percent and 53.4 percent, respectively.^[2][6] Differences in utility and coverage complicate direct ranking, but the nonzero rates support Meta's recommendation to add deployment-specific safeguards.

Limitations

Muse Glimmer can produce inaccurate, biased, or objectionable responses and can fail during long or unfamiliar multi-step tasks. It was not evaluated in every language represented in training. Image understanding does not imply audio support or purpose-built video reasoning, and lower-precision inference may affect edge cases.^[2]

The local-agent framing also needs a hardware qualification. Meta released compact quantizations that fit its 24 GB and 32 GB target envelopes under stated assumptions, but long context and full multimodal or speculative components consume additional memory. Performance and fit should be measured on the intended runtime and workload rather than inferred from a GGUF filename.

Finally, benchmark and safety claims available at release came primarily from Meta. The methodology report is more informative than an unqualified leaderboard because it identifies score provenance, judges, trial counts, and harness differences, but it remains publisher evaluation.^[6] Independent release coverage establishes that Muse Glimmer is a real, publicly downloadable model. It does not convert Meta's measured results into an independently reproduced consensus.^[8]

References

1. ^Meta AI Research, *Introducing Muse Glimmer: An Open Agentic Model That Runs on Your Device*, August 10, 2026. research.meta.ai/...use-glimmer-open-agentic-model
2. ^Meta Superintelligence Labs, *Muse Glimmer Model Card*, accessed August 15, 2026. huggingface.co/...Muse-Glimmer-30B
3. ^Meta, *Muse Glimmer released configuration*, accessed August 15, 2026. huggingface.co/...config.json
4. ^Hugging Face, *API metadata for meta-models/Muse-Glimmer-30B*, accessed August 15, 2026. huggingface.co/...Muse-Glimmer-30B
5. ^Meta, *Muse-Glimmer-30B-GGUF repository*, accessed August 15, 2026. huggingface.co/...Muse-Glimmer-30B-GGUF
6. ^Meta AI Research, *Muse Glimmer Evaluation Methodology*, August 2026. research.meta.ai/...muse-glimmer-methodology
7. ^Meta for Developers, *Muse Glimmer*, accessed August 15, 2026. developer.meta.com/...muse-glimmer
8. ^Kelvin Chan and Barbara Ortutay, *Zuckerberg outlines Meta's ambitions for world-changing AI technology*, Associated Press, August 10, 2026. apnews.com/...nce-df8a4e7d7825470d09e8090367457c2c
9. ^Jian Chen, Yesheng Liang, and Zhijian Liu, *DFlash: Block Diffusion for Flash Speculative Decoding*, arXiv:2602.06036, accepted at ICML 2026, revised May 28, 2026. arxiv.org/...2602.06036
10. ^Daniel Bolya et al., *Perception Encoder: The best visual embeddings are not at the output of the network*, arXiv:2504.13181, revised April 28, 2025. arxiv.org/...2504.13181
11. ^Meta, *Muse Glimmer Usage Policy*, accessed August 15, 2026. huggingface.co/...USAGE_POLICY.md