Muse Glimmer on Ollama: Meta's 30B open agent model explained
On August 10 2026 Meta Superintelligence Labs released its first open model: Muse Glimmer. It is a 30-billion-parameter multimodal model purpose-built for agentic workloads that run locally. It is distilled from Muse Spark, released under Apache 2.0 and available on Ollama from day one.
Where most recent open AI models chase general chat quality, Muse Glimmer has a narrower mission: power coding agents such as Claude Code, Codex, Pi and OpenCode plus long-running personal assistants such as OpenClaw and Hermes — entirely on your own hardware with no cloud and no per-token billing.
This guide covers the model in depth: architecture, reasoning modes, hardware needs, Ollama setup, benchmarks and a detailed comparison with Gemma 4, Qwen 3.6 and other recent open models so you can decide if Muse Glimmer belongs on your machine.
What Muse Glimmer actually is
| Property | Value |
|---|---|
| Release date | August 10 2026 by Meta Superintelligence Labs |
| Parameters | About 30B |
| Context window | 128K tokens by default with a 131072 token maximum |
| Languages | More than 100 languages |
It is a dense model, not a mixture-of-experts. Every one of the 30B parameters is active on every token. That brings predictable latency, reliable instruction following and fewer routing failure modes across long tool chains. The price is higher compute per token than a sparse MoE such as Qwen 3.6 35B-A3B where only about 3B parameters are active per token.
Why it is built for agents
Muse Glimmer was trained and evaluated for end-to-end agentic task completion rather than single-turn Q and A. Five capabilities define it:
Reliable tool use. The model invokes tools with precise schemas across extended workflows. Note the output format: Muse Glimmer emits channel-scoped reasoning plus XML-style ATEM tool calls rather than plain JSON. Frameworks need the dedicated muse_glimmer tool-call and reasoning parsers in vLLM. On Ollama this is handled for you through ollama launch integrations.
Multi-step reasoning. Muse Glimmer chains reasoning over long horizons and sustains coherent plans across workflows that span dozens of tool calls.
Failure recovery. When a tool call fails or returns an unexpected result the model diagnoses the error and retries instead of halting. This is one of the behaviours Meta tuned explicitly for always-on agents.
Framework compatibility. It works across OpenClaw, Hermes Agent and other orchestration patterns without framework-specific fine-tuning.
Controllable effort. Four reasoning strengths are exposed: low, medium, high and xhigh. Use high or xhigh for complex coding and agentic tasks. Use low or medium when speed matters more than depth.
If you only want everyday chat, toggle visible reasoning off with --think=false on Ollama: see our guide to disabling thinking mode.
Multimodal input rounds out the agent story. The 1.8B perception encoder reads screenshots, charts, receipts, mockups and documents alongside text. Practical uses include building a site from a drawing, driving computer-use agents from screenshots and reading scanned documents locally.
Running it with Ollama
You need Ollama 0.31 or newer for the best experience, especially on Apple Silicon where DFlash and image input landed recently. The basic flow is familiar:
# Default dense build, about 18 GB
ollama run muse-glimmer
Hardware requirements
Muse Glimmer is designed to fit a single consumer GPU, but "fits" depends on quantization and context length. As a rule of thumb for the default Ollama quants:
| Your machine | What to expect |
|---|---|
| 8 GB VRAM such as RTX 3070 | Possible only with community ultra-low-bit GGUF builds such as UD-IQ2_XXS at about 11 GB download plus CPU and RAM offloading. Expect a reduced default context near 8K and modest speed. |
| 16 GB VRAM or unified memory | Tight for the official 18 GB build. Prefer the MLX build on a 24 GB Mac or a smaller model such as Gemma 4 12B or Qwen 3.6 with lower quants. |
| 24 GB VRAM such as RTX 4090 | The sweet spot. The 18 GB Q4 build fits with room for a large num_ctx. |
| 32 GB VRAM such as RTX 5090 or 32 GB Mac | Comfortable. Run the MLX or NVFP4 build at full 128K context with headroom for agents that hold many tool results. |
| Workstation such as DGX Spark | Ideal host for always-on agents at full context with speculative decoding through the 5 GB DFlash draft head. |
Two levers control the footprint exactly as described in our complete Ollama guide: parameter count and quantization. A third lever matters a lot here: num_ctx. The 128K window is generous for agents that accumulate tool outputs, but it costs memory through the KV cache. If the model spills from VRAM to system RAM and slows down, lower num_ctx to 32768 or 8192 before blaming the model. Muse Glimmer is unusually forgiving here thanks to its 52 KiB per token cache.
An agent that reads files, runs commands and keeps terminal output quickly fills 32K tokens. Start with a large num_ctx on a 24 GB card and reduce it only if you see spilling in ollama ps. Pair this with the memory rules from our coding model guide.
Apple Silicon, MLX and DFlash
On Macs, use the MLX tag: ollama pull muse-glimmer:30b-mlx. It is optimized for the Ollama MLX engine and gets you two benefits: DFlash speculative decoding (a small draft head proposes up to 15 tokens at once, verified in a single pass) for 1.5x to 1.8x faster generation with zero configuration, and native image input in the MLX path for screenshot-driven agents and low-latency mockup-to-code workflows.
Muse Glimmer is heavier than the edge models in our Gemma 4 on Apple Silicon guide (30 to 50 tokens per second for E2B and E4B on M4 hardware, about 25 tokens per second for the 12B on 16 GB machines). Expect the best speeds on 24 GB Macs and above. The MLX build remains the recommended tag on any M-series Mac.
What NVIDIA reports
NVIDIA positions Muse Glimmer as an optimal balance for local agents: large enough for complex multi-step reasoning yet small enough to avoid model sharding or CPU offloading on a single GPU. Reported figures on Blackwell Ultra exceed 20000 tokens per second per GPU at BF16 and NVF4 precision with high concurrency sustained without MoE routing overhead.
Benchmarks in context
Meta compared Muse Glimmer in high-reasoning mode against Gemma 4 31B in thinking mode and Qwen 3.6 27B in thinking mode. Independent tracking through Artificial Analysis tells a broadly similar story with small ranking differences. Here are the headline numbers from Meta evaluation:
| Category | Benchmark | Muse Glimmer 30B | Gemma 4 31B | Qwen 3.6 27B |
|---|---|---|---|---|
| General agentic | MCP Atlas public | 75.5 | 54.2 | 62.5 |
| General agentic | DeepSearch QA | 74.6 | 61.7 | 71.1 |
| General agentic | Tau3-Banking | 23.5 | 15.1 | 16.7 |
| General agentic | WildClawBench | 47.6 | 37.6 | 43.2 |
| General agentic | Gaia2 | 43.3 | 36.4 | 40.0 |
| Agentic coding | SWE-Bench Pro | 51.2 | 36.9 | 50.2 |
| Agentic coding | SWE-Bench Verified | 76.0 | 66.6 | 77.2 |
| Agentic coding | TerminalBench 2.1 | 51.7 | 43.4 | 60.7 |
| Agentic coding | SciCode | 43.6 | 43.4 | 39.8 |
| Multimodal | Charxiv reasoning | 78.8 | 77.7 | 78.4 |
| Multimodal | ScreenSpot Pro | 75.4 | 75.9 | 76.1 |
| Multimodal | OmniDocBench 1.5 | 75.8 | 72.5 | 77.8 |
| Reasoning | AIME 2026 | 94.7 | 89.2 | 94.1 |
| Reasoning | IFBench | 77.0 | 76.0 | 70.8 |
| Reasoning | AA-LCR long context | 80.0 | 68.3 | 73.3 |
| Reasoning | Beam128K | 65.1 | 58.2 | 63.0 |
Three takeaways. Muse Glimmer dominates the agent benchmarks (MCP Atlas by 13 points over Qwen and 21 over Gemma, plus clear leads on DeepSearch QA, Tau3-Banking, WildClawBench and Gaia2). On coding it runs neck and neck with Qwen: Muse Glimmer leads SWE-Bench Pro and SciCode, Qwen leads SWE-Bench Verified and TerminalBench, and Gemma 4 trails both. Multimodal and long-context reasoning favour Muse Glimmer, while Qwen keeps a small edge on OSWorld and GDPVal-style server tasks.
Comparison with recent open models
Side-by-side summary
| Model | Params and arch | Context | License | Ollama pull size | Best for |
|---|---|---|---|---|---|
| Muse Glimmer 30B | 29.6B dense plus 1.8B vision | 128K | Apache 2.0 | About 18 GB | Always-on local agents, tool loops and failure recovery |
| Gemma 4 31B | 30.7B dense | 256K | Apache 2.0 | About 19 GB | Multilingual chat, multimodal breadth and human preference |
| Gemma 4 26B-A4B | 25.2B MoE with 3.8B active | 256K | Apache 2.0 | About 15 GB | Best quality per GB on laptops and mid-range GPUs |
| Gemma 4 12B | 11.95B unified | 256K | Apache 2.0 | About 7 GB | Single model for chat plus light coding on 16 GB machines |
| Qwen 3.6 27B | 27B dense | 262K | Apache 2.0 | About 16 GB | Pure code generation and reasoning on 24 GB cards |
| Qwen 3.6 35B-A3B | 35B MoE with 3B active | 262K | Apache 2.0 | About 21 GB | Efficient agentic coding at low per-token cost |
| DeepSeek-R1 distill 32B | 32B dense | 128K | MIT | About 20 GB | Math, step-by-step debugging and logic-heavy tasks |
| GLM-5.2 | 744B MoE | 1M | MIT | Server only | Long-horizon software engineering agents |
| Kimi K3 | 1000B MoE | 1M | Bespoke open-weight | Server only | Marathon agents and top open coding scores |
Limits and honest trade-offs
Muse Glimmer is strong but not universal. Its 128K context trails Gemma 4 at 256K, Qwen 3.6 at 262K and frontier models at 1M. Repository-scale analysis across hundreds of files still favours larger windows. Its dense design costs more per token than MoE rivals with similar quality. The XML-style ATEM tool format needs parser support outside Ollama, while Qwen and Gemma stick to more conventional JSON tool schemas. Vision covers text plus image input only, with no audio or video tracks such as Gemma E-models provide. Safety benchmarks show solid utility but confirm that any agent with shell access needs sandboxing, timeouts and human review for destructive actions.
None of these are disqualifiers. They define where Muse Glimmer fits: the best open agent brain that still runs on one good consumer GPU with a clean Apache 2.0 license and a tiny KV cache that keeps long tool sessions practical.
Meta is back in open weights after the long Llama silence. Muse Glimmer is not the largest model of 2026 and it does not try to be. It is the most agent-focused 30B model you can self-host today — and that makes it one of the most useful Ollama pulls of the year.