Overview
Muse Glimmer is Meta's August 10, 2026 open-weight release in the Muse model family. Meta introduced it as a 30-billion-parameter model optimized for local, always-on AI agent workflows, shifting the Muse line from the hosted Muse Spark assistant experience toward a model that developers can evaluate for local or self-managed use.
The launch is important because it changes the practical question around Muse. Spark was mostly a Meta-platform assistant; Glimmer is positioned as a compact open-weight agent model competing with other sub-40B reasoning models such as Gemma4-31B Thinking and Qwen3.6-27B Thinking. Meta has published a benchmark comparison in the official AI at Meta launch post, but it has not yet surfaced a first-party model card, license page, download repository, or hardware guide that confirms exactly how teams can deploy it in production.
What's New
Open-Weight Muse Model
Muse Glimmer is the first Muse release Meta has publicly framed as open-weight. That matters for developers who want to inspect, host, benchmark, quantize, or adapt the model outside Meta's consumer apps. It also partially answers a limitation of Muse Spark: Spark was powerful inside Meta AI, WhatsApp, Instagram, Facebook, and Messenger, but it did not give builders a public weight release or broad self-hosting path.
Open-weight should not be treated as the same thing as a confirmed open-source license. As of this release page, Meta's official X announcement confirms the open-weight positioning and 30B parameter size, but not the license terms, commercial-use permissions, acceptable-use policy, training disclosure, model-card details, or redistribution rules. Teams should verify the final license before using Glimmer in commercial products.
Local, Always-On Agent Workflows
Meta describes Muse Glimmer as optimized for local, always-on agent workflows. In practice, that points to workloads where latency, privacy, persistence, and cost predictability matter:
- Desktop or workstation agents that stay active across files, browser state, and local tools
- Coding assistants that need repeated tool calls without sending every step to a hosted frontier model
- Personal productivity agents where local context and lower marginal cost matter
- Research prototypes comparing open-weight models for continuous background automation
The official post does not yet confirm runtime integrations such as Ollama, llama.cpp, vLLM, SGLang, MLX, or Hugging Face Transformers. It also does not publish minimum RAM or VRAM requirements. For now, buyers should treat "local" as a product direction, not as proof that Glimmer will run well on ordinary laptops without quantization.
Benchmark Positioning
Meta's benchmark image compares Muse Glimmer-30B High Reasoning with Gemma4-31B Thinking Mode and Qwen3.6-27B Thinking Mode. Glimmer leads several agentic and general capability rows in Meta's chart, including MCP Atlas, DeepSearch QA, T3-Banking, WildClawBench, GAIA2, SWE-Bench Pro, SciCode, Charxiv Reasoning, IFBench, AIME 2026, AA-LCR, and Beam 128K.
The same chart also shows Qwen3.6-27B leading on GDPval-AA, SkillsBench with skills, OSWorld-Verified, SWE-Bench Verified, TerminalBench 2.1, ScreenSpot Pro, OmniDocBench v1.5, and MMMU Pro, while Gemma4-31B leads GPQA Diamond and Humanity's Last Exam in the reported comparison. That makes Glimmer strongest as a local-agent candidate rather than a universal benchmark winner.
From Muse Spark to Glimmer
Muse Spark was Meta's first Muse model from Meta Superintelligence Labs, with a hosted assistant focus: multimodal perception, Contemplating mode, visual coding, health-informed responses, and Meta-platform integration. Glimmer is narrower but more developer-relevant: it emphasizes local, always-on agent workflows and open weights.
That difference is useful for evaluation. If you want a free assistant inside Meta's apps, Spark remains the more direct user-facing product. If you want to test a Meta model in self-managed agent infrastructure, Glimmer is the release to watch once model files and license terms are available.
Availability & Access
Meta has publicly announced Muse Glimmer through the official AI at Meta X account. The announcement confirms:
| Item | Status |
|---|---|
| Model name | Muse Glimmer |
| Parameter size | 30B |
| Weight availability | Announced as open-weight |
| Target use case | Local, always-on agent workflows |
| Official benchmark chart | Published in the launch post |
| Model repository | Not confirmed in public sources checked for this page |
| License | Not confirmed in public sources checked for this page |
| Hardware requirements | Not confirmed in public sources checked for this page |
| Public API | Not confirmed in public sources checked for this page |
System Requirements & Limitations
No official deployment requirements are available yet. A 30B model commonly needs quantization or high-memory hardware for comfortable local use, but the exact practical requirement depends on weight format, precision, context length, KV cache behavior, and inference runtime. Do not assume consumer-laptop readiness until Meta publishes the model files and supported serving guidance.
Pricing & Plans
Muse Glimmer is an open-weight model announcement, so the software price is currently best treated as free to access once weights are available. The practical cost will come from deployment:
| Cost Area | What to Budget For |
|---|---|
| Local inference | GPU/CPU hardware, RAM/VRAM, storage, power, and maintenance |
| Hosted inference | Third-party inference provider pricing if providers add Glimmer endpoints |
| Commercial use | License review and compliance work once official terms are available |
| Evaluation | Benchmarking against models already in your stack, especially Qwen, Gemma, Llama, DeepSeek, and Kimi variants |
Meta has not announced API pricing, hosted usage tiers, or managed inference for Glimmer. Teams that need production-grade SLAs should wait for official deployment channels or third-party providers before relying on it operationally.
Best For
- Developers evaluating open-weight models for local desktop agents and background automation
- Teams comparing compact reasoning models around the 27B-31B range, especially against Gemma and Qwen thinking variants
- AI infrastructure teams that want to test Meta's post-Spark open-weight direction without waiting for a larger Llama-class release
- Privacy-conscious builders who prefer self-managed inference once license and deployment details are available
- Researchers tracking agentic benchmark performance across MCP Atlas, DeepSearch QA, SWE-Bench Pro, TerminalBench, and OSWorld-style tasks
FAQ
What is Muse Glimmer?
Muse Glimmer is a 30B open-weight model announced by AI at Meta on August 10, 2026. Meta positions it for local, always-on agent workflows rather than only hosted assistant use inside Meta's apps.
Is Muse Glimmer open source?
Not confirmed. Meta's official post calls Glimmer open-weight, but a license page or model card was not found during this review. Until those terms are published, it is safer to say open-weight rather than open source.
Can I run Muse Glimmer locally?
Meta describes Glimmer as optimized for local workflows, but official runtime and hardware requirements are not yet published. Wait for the model repository, quantization formats, and serving guidance before planning production local deployment.
How does Muse Glimmer compare with Muse Spark?
Muse Spark is the Meta-hosted assistant model used across Meta AI and Meta apps. Muse Glimmer is a 30B open-weight model aimed at local agent workflows. Spark is better framed as a consumer assistant experience; Glimmer is better framed as a developer-facing model release.
What benchmarks did Meta publish for Muse Glimmer?
Meta's launch chart compares Muse Glimmer-30B with Gemma4-31B Thinking and Qwen3.6-27B Thinking across agentic, coding, multimodal, safety, instruction-following, and reasoning benchmarks. Glimmer leads several rows, including MCP Atlas, DeepSearch QA, SWE-Bench Pro, IFBench, AIME 2026, and Beam 128K in Meta's reported chart.
