--- source: 'https://howaiworks.ai/models/muse-glimmer' section: models title: Muse Glimmer description: >- Meta's Apache 2.0 open-weight 30B multimodal agent model, distilled from Muse Spark and built to run offline on a single 24 GB consumer GPU. tags: - Muse - Meta - Open Weights - Apache 2.0 - Agentic AI - Multimodal - Language Model - Local AI - On-Device AI - Latest category: Language Models developer: Meta developerWebsite: 'https://ai.meta.com/' modelType: Open-Weights Multimodal Model releaseDate: '2026-08-10' lastUpdated: '2026-09-06' isReviewed: false license: Apache 2.0 contextWindow: 128K tokens knowledgeCutoff: January 2026 officialWebsite: 'https://developer.meta.com/ai/models/muse-glimmer/' docsPage: 'https://ai.developer.meta.com/docs/muse-glimmer' pricingPage: 'https://openrouter.ai/meta/muse-glimmer-30b' --- # Muse Glimmer > Meta's Apache 2.0 open-weight 30B multimodal agent model, distilled from Muse Spark and built to run offline on a single 24 GB consumer GPU. ## Overview Muse Glimmer is a **30-billion-parameter open-weight multimodal model** released by **Meta on 10 August 2026** under the **Apache 2.0** licence. It is built for one job: running an agent locally, on hardware someone already owns, with no API key and no per-token bill. It is not a small sibling of [Muse Spark](https://howaiworks.ai/models/muse-spark) in the usual sense. Meta built it by **distilling Muse Spark's outputs**, then post-training with policy distillation and reinforcement learning against agentic trajectories. The result is a dense model that inherits Muse Spark's tool-use behaviour at a size that fits on a desk. The strategic reading matters more than the benchmark table. Muse Spark, launched in April 2026, was **closed-weights** — a break from the open lineage that defined [Llama](https://howaiworks.ai/models/llama). Muse Glimmer reverses that, and it does so with the most permissive licence Meta has ever applied to a model: not a community licence with a monthly-active-user threshold, not an OpenMDW-style research licence, but plain Apache 2.0. Meta's open-weight programme did not end with Llama 4. One caveat on that reversal, because it is widely misreported. On launch day Mark Zuckerberg also said Meta would "soon" release the weights for **Muse Spark 1.2**, and repeated the promise on 2 September 2026 without naming a date. **Those weights have not shipped.** Meta's flagship page has since moved on to Muse Spark 1.3. As of this writing, Muse Glimmer is the only downloadable Muse model. ## Capabilities - **Agentic tool use over long trajectories** — the capability the model was built and post-trained for. Meta's framing is that tool calling "holds up over long sequences" rather than degrading after a handful of calls, which is the usual failure mode for models of this size. - **Failure recovery** — when a tool call errors or returns something unexpected, the model is trained to diagnose and retry rather than hallucinate a result and continue. This is the difference between an agent you can leave running and one you have to babysit. - **Interleaved text and image input** — a dedicated perception encoder lets an agent read a screenshot, a chart or a scanned document as part of a reasoning step, not as a separate captioning call. Output is text only. - **Agentic coding** — multi-file edits, terminal work and compatibility with existing coding harnesses. Meta ships its own terminal agent, Muse Code, though that was released alongside Muse Spark 1.2 rather than for Glimmer. - **Step-by-step reasoning before answering** — the model reasons before it responds; Meta's prompting guide treats this as default behaviour rather than an optional mode. - **Block speculative decoding** — a separately released drafter proposes whole blocks of tokens which the main model verifies in parallel. Meta reports 3.1x generation speedup on an RTX 5090, 1.8x on an M5 Max and 1.5x on an M4 Max. This is what makes a 30B dense model tolerable inside an agent loop on consumer silicon. - **Multilingual** — trained on data from more than 100 languages. Meta has published no per-language quality breakdown. ## Technical Specifications - **Architecture**: dense causal transformer — **no [Mixture-of-Experts](https://howaiworks.ai/glossary/mixture-of-experts) routing**. All parameters activate on every forward pass. - **Parameters**: ~29.6B total, including the vision encoder - **Layers**: 52, with an alternating local/global attention pattern - **Hidden dimension**: 6,656; SwiGLU FFN with intermediate dimension 19,968 - **Attention**: 32 query heads / 2 KV heads — a 16:1 [grouped-query attention](https://howaiworks.ai/glossary/grouped-query-attention) ratio, which is what keeps the KV cache small enough for a 24 GB card at long context - **Position encoding**: RoPE, θ = 500,000 - **Perception encoder**: ~1.8B-parameter [ViT](https://howaiworks.ai/glossary/vision-transformer)-G/14, 50 layers, width 1536, patch size 14; up to **4,096 visual tokens per image** - **[Context window](https://howaiworks.ai/glossary/context-window)**: **131,072 tokens (128K)**. Meta says longer is supported but names no ceiling. - **Maximum output**: **not published by Meta.** - **Knowledge cutoff**: **4 January 2026** - **Modalities**: text + image in, text out - **Languages**: 100+ - **Weights**: `meta-models/Muse-Glimmer-30B` (BF16 and 4-bit) and `meta-models/Muse-Glimmer-30B-GGUF` on Hugging Face ### Quantization and the real VRAM floor The "runs on 24 GB" headline is a claim about a specific 4-bit build, not about the model as shipped. Meta publishes three: | Build | File size | VRAM, text only | + vision encoder | + drafter | Meta's stated accuracy loss | |---|---|---|---|---|---| | BF16 full precision | ~60 GB | 55-64 GB | — | — | baseline | | K-Quant-Dynamic (Q4_K_XL) | 19.7 GB | ~20 GB | ~22 GB | ~23 GB | 0.2% | | K-Quant-17GB (Q4_K_M) | 16.8 GB | ~17 GB | ~19 GB | ~20 GB | 1.0% | Read the columns, not the headline. A 24 GB RTX 3090 or 4090 running the K-Quant-17GB build with vision *and* the speculative drafter loaded is sitting at roughly 20 GB before a single token of context is cached — workable, but not spacious, and nowhere near the full 128K window. The K-Quant-Dynamic build targets 32 GB machines (RTX 5090, M5 Max) and gives up almost nothing to full precision. Both degradation figures are [quantization](https://howaiworks.ai/glossary/quantization) claims measured by Meta across 15 benchmarks; nobody has independently reproduced them. A **DFlash drafter** ships separately for [speculative decoding](https://howaiworks.ai/glossary/speculative-decoding). Meta suggests 5 drafted tokens per step as a baseline, 8-10 for code, 3-4 for open-ended generation. ## Use Cases - **Always-on local agents** — background agents that watch a directory, a repository or a queue and act without a network round trip. The economics only work when inference is free, which is the whole argument for a local model. - **Coding agents on private repositories** — where the code cannot leave the machine for contractual or regulatory reasons and a hosted frontier model is simply not an option. - **Document and screenshot workflows** — the perception encoder makes it viable to route scanned invoices, charts and UI screenshots through the same agent that handles the text, instead of stitching an OCR service into the loop. - **Air-gapped deployment** — no API key, no telemetry, no dependency on a vendor's uptime or on a model that may be deprecated out from under you. Apache 2.0 means this is a genuine option, not a licensing negotiation. - **Cost-floor batch work** — high-volume classification, extraction or triage where per-token pricing dominates and the marginal quality of a frontier model does not repay it. - **A base for [supervised fine-tuning](https://howaiworks.ai/glossary/supervised-fine-tuning)** — day-zero TRL support for SFT, [LoRA](https://howaiworks.ai/glossary/lora) and AsyncGRPO. LoRA is reported to fit a single 80 GB H100; full SFT wants 8x80 GB with FSDP or ZeRO-3, so "runs on a consumer GPU" applies to inference only. Where it is the wrong choice: anything where a frontier model's ceiling is the point. Glimmer is a distillation of Muse Spark and Meta says so plainly on the model card — it is "generally less capable than Muse Spark." ## Performance / Benchmarks All figures below are **Meta-run**, published on Meta's own model page alongside the 10 August launch, comparing against **Gemma4-31B** and **Qwen3.6-27B**. No independent reproduction exists at the time of publication. | Benchmark | Muse Glimmer | Gemma4-31B | Qwen3.6-27B | |---|---|---|---| | MCP Atlas (tool orchestration) | **75.5** | 54.2 | 62.5 | | DeepSearch QA | **74.6** | 61.7 | 71.1 | | GAIA2 | **43.3** | 36.4 | 40.0 | | WildClawBench | **47.6** | 37.6 | 43.2 | | τ³-Banking | **23.5** | 15.1 | 16.7 | | SkillsBench | 44.3 | 32.4 | **46.6** | | OSWorld-Verified | 65.9 | 58.5 | **75.6** | | GDPval-AA | 953 | 811 | **1141** | | SWE-Bench Pro | **51.2** | 36.9 | 50.2 | | SWE-Bench Verified | 76.0 | 66.6 | **77.2** | | TerminalBench 2.1 | 51.7 | 43.4 | **60.7** | | SciCode | **43.6** | 43.4 | 39.8 | | MMMU Pro | 74 | 73 | **75** | | Charxiv Reasoning | **78.8** | 77.7 | 78.4 | | ScreenSpot Pro | 75.4 | 75.9 | **76.1** | | OmniDocBench v1.5 | 75.8 | 72.5 | **77.8** | | AIME 2026 | **94.7** | 89.2 | 94.1 | | GPQA Diamond | 83.5 | **85.7** | 84.2 | | IFBench | **77.0** | 76.0 | 70.8 | | AA-LCR | **80.0** | 68.3 | 73.3 | | Beam 128K | **65.1** | 58.2 | 63.0 | The honest summary is narrower than the marketing. Against Gemma4-31B, Muse Glimmer wins nearly everything. Against Qwen3.6-27B it wins **tool orchestration, agentic search, long-context recall and instruction following**, and it loses **computer use (OSWorld-Verified), terminal work (TerminalBench 2.1), SkillsBench and GDPval-AA** — several of them by wide margins. Multimodal results are a wash across all three. If your agent's inner loop is MCP tool calls and long-horizon planning, the table favours Glimmer. If it is driving a GUI or a shell, it does not. ## Limitations - **The 24 GB claim assumes 4-bit.** At BF16 this is a 55-64 GB model. Everything the marketing says about consumer hardware is downstream of quantization, and the 24 GB build is the one Meta measures at 1.0% degradation rather than 0.2%. - **24 GB is a tight fit, not a comfortable one.** Vision encoder plus drafter puts you near 20 GB before context. Long-context work on a 24 GB card will force trade-offs between window size, vision and speculative decoding. - **Dense, so it is slow for its size.** No MoE routing means all ~29.6B parameters run every forward pass. Speculative decoding is the mitigation, not an architectural advantage. - **Beaten on computer use and terminal work** by Qwen3.6-27B on Meta's own numbers — a notable admission in a vendor-published table, and disqualifying if that is your workload. - **All benchmarks are vendor-run.** Meta Superintelligence Labs published them; nobody has reproduced them independently at the time of publication. - **No published maximum output length**, and no named ceiling above the 128K default context. - **Text output only.** No audio input, no image generation, no speech. - **Materially weaker than Muse Spark**, by Meta's own model card. A distilled 30B model is not a frontier model with a smaller footprint. - **Standard open-model caveats apply** — Meta's card warns it "may produce inaccurate, biased, or objectionable responses" and can "still make errors in multi-step reasoning, particularly in novel scenarios." An always-on agent amplifies the consequences of both. - **Fine-tuning is not a consumer-hardware activity.** Inference fits a 24 GB card; LoRA wants an 80 GB H100 and full SFT wants eight of them. ## Pricing & Access **The weights are free and the licence is Apache 2.0.** There is no acceptable-use addendum, no user threshold, no attribution requirement beyond what Apache 2.0 itself asks, and no restriction on commercial deployment or on distributing derivatives. That covers the BF16 weights, both quantized builds, the DFlash drafter and the perception encoder. **Meta operates no first-party paid endpoint for Muse Glimmer**, so there is no vendor rate card and no price this page can put in its header. Hosted access comes from third parties, who set their own prices: - **OpenRouter** aggregates Phala, DeepInfra, Fireworks and Together. Input sits around **$0.30 per 1M tokens** across providers; output ranges from **$1.10 to $1.50 per 1M tokens** depending on which one serves the request, with cache reads around $0.04 per 1M. Those are third-party prices, not Meta's, and they move. - **Together AI** and **Fireworks AI** serve it directly as well. For most readers arriving at this page, the relevant price is zero: download the GGUF, run it in [Ollama](https://howaiworks.ai/ai-tools/ollama) or [LM Studio](https://howaiworks.ai/ai-tools/lm-studio), and the only cost is electricity. ## Ecosystem & Tools Meta shipped day-zero integrations rather than weights and a shrug: - **Local runtimes** — [Ollama](https://howaiworks.ai/ai-tools/ollama), [LM Studio](https://howaiworks.ai/ai-tools/lm-studio) and Unsloth all package the model; llama.cpp consumes the official GGUF k-quants directly. Unsloth publishes its own optimised quant set alongside Meta's. - **On-device** — ExecuTorch builds for mobile and embedded targets. - **Server-scale serving** — vLLM and SGLang, for anyone running it as shared infrastructure rather than on a workstation. - **Hugging Face** — `transformers` support at launch, plus Inference Endpoints. The GGUF repository carries 161 quantized variants between Meta's and the community's. - **Fine-tuning** — TRL, covering SFT, LoRA and AsyncGRPO. - **Hosted APIs** — Together AI, Fireworks AI and [OpenRouter](https://howaiworks.ai/ai-tools/openrouter). - **Related catalogue pages** — [Muse Spark 1.3](https://howaiworks.ai/models/muse-spark) is the frontier model Glimmer was distilled from; [Llama 4](https://howaiworks.ai/models/llama) is Meta's previous open-weight family; [Gemma 4](https://howaiworks.ai/models/gemma) and [Nemotron 3 Ultra](https://howaiworks.ai/models/nemotron) are the other major open-weight releases a self-hoster is weighing this against. ## Community & Resources - [Muse Glimmer model page](https://developer.meta.com/ai/models/muse-glimmer/) — Meta's specifications and the full benchmark table - [Muse Glimmer: an open-weight model built for local agents](https://developer.meta.com/ai/resources/blog/build-with-muse-glimmer/) — Meta's launch write-up on the AI Developers blog - [Muse Glimmer documentation](https://ai.developer.meta.com/docs/muse-glimmer) — deployment guides for vLLM, SGLang, llama.cpp and ExecuTorch, plus the prompting and quantization pages - [meta-models/Muse-Glimmer-30B](https://huggingface.co/meta-models/Muse-Glimmer-30B) — the model card, BF16 and 4-bit weights - [meta-models/Muse-Glimmer-30B-GGUF](https://huggingface.co/meta-models/Muse-Glimmer-30B-GGUF) — llama.cpp builds, including the 24 GB K-Quant-17GB - [Meta is back with Muse Glimmer](https://huggingface.co/blog/muse-glimmer) — Hugging Face's integration write-up ## Frequently Asked Questions ### What licence is Muse Glimmer under? Apache 2.0 — the permissive OSI licence, with no acceptable-use addendum, no monthly-active-user threshold and no naming requirement. Meta describes it as "the most permissive license we've used for an open model." It covers the full-precision weights, both quantized builds, the speculative-decoding drafter and the perception encoder. Commercial use is unrestricted. ### Does Muse Glimmer really run on a 24 GB GPU? Yes, but only in the 4-bit K-Quant-17GB build. That file is 16.8 GB and needs roughly 17 GB of VRAM for text, ~19 GB with the vision encoder loaded and ~20 GB with the speculative drafter as well — which fits a 24 GB card such as an RTX 3090 or 4090 with little headroom for long contexts. BF16 needs 55-64 GB. The 24 GB claim is a quantization claim, not a claim about the model as shipped. ### How much accuracy does the 4-bit build lose? Meta reports 1.0% average degradation across 15 benchmarks for the 24 GB K-Quant-17GB build, and 0.2% for the larger K-Quant-Dynamic build aimed at 32 GB systems. Both figures are Meta's own measurements and have not been independently reproduced at the time of publication. ### What is the context window of Muse Glimmer? 128K tokens by default — 131,072, the figure shown on the model card and on hosted endpoints. Meta's documentation says longer contexts are supported but does not name a ceiling, and Meta has not published a maximum output length. ### Where do I download the weights? Hugging Face, at meta-models/Muse-Glimmer-30B for the BF16 and 4-bit checkpoints and meta-models/Muse-Glimmer-30B-GGUF for the llama.cpp builds. Ollama, LM Studio and Unsloth redistribute their own packagings of the same weights. ### Is Muse Glimmer multimodal? Input-side, yes. It carries a dedicated ~1.8B-parameter ViT-G/14 perception encoder and accepts interleaved text and images, up to 4,096 visual tokens per image. Output is text only. There is no audio input and no image generation. ### Did Meta also open-weight Muse Spark 1.2? Not yet. On 10 August 2026 Mark Zuckerberg said Meta would "soon" release the weights for Muse Spark 1.2, and repeated the promise on 2 September 2026 without a date. As of 6 September 2026 no Muse Spark weights have shipped, and Meta's own model page has since moved to Muse Spark 1.3. Muse Glimmer is currently the only Muse model you can download. ### How does Muse Glimmer compare to Gemma 4 31B and Qwen3.6-27B? On Meta's own comparison it leads both on tool orchestration, agentic search and long-context recall, and beats Gemma 4 on essentially everything. Against Qwen3.6-27B the picture is mixed: Muse Glimmer wins SWE-Bench Pro and AIME 2026 but loses OSWorld-Verified, TerminalBench 2.1, SkillsBench and GDPval-AA. All figures are vendor-run. ### Is Muse Glimmer a Mixture-of-Experts model? No. It is a dense causal transformer — roughly 29.6B parameters across 52 layers, all of them active on every forward pass. That is the reason its VRAM footprint is predictable and the reason it is slower per token than a sparse model of the same nominal size. ### What does Muse Glimmer cost to use? Nothing to self-host. Meta operates no first-party paid endpoint for it, so there is no vendor rate card. Third-party hosts price it independently: on OpenRouter it is around $0.30 per 1M input tokens with output between $1.10 and $1.50 depending on which provider serves the request. ## Related ### Related models - [Muse Spark 1.3](https://howaiworks.ai/models/muse-spark) - [Gemma 4](https://howaiworks.ai/models/gemma) - [Llama 4](https://howaiworks.ai/models/llama) - [Nemotron 3 Ultra](https://howaiworks.ai/models/nemotron) - [GLM-5.3](https://howaiworks.ai/models/glm) - [Mistral Medium 3.5](https://howaiworks.ai/models/mistral-medium) ### Related tools - [Ollama](https://howaiworks.ai/ai-tools/ollama) - [LM Studio](https://howaiworks.ai/ai-tools/lm-studio) - [OpenRouter](https://howaiworks.ai/ai-tools/openrouter) --- Source: https://howaiworks.ai/models/muse-glimmer — HowAIWorks.ai