What it is
Muse Glimmer is an open-weight model from Meta Superintelligence Labs, released on August 10, 2026. It was Meta's first openly licensed model since Muse Spark replaced the Llama line in April, and it is distilled from Muse Spark. It is a dense model of about 29.6 billion parameters, including a 1.8-billion-parameter image encoder, and reads text and images. The context is 131,072 tokens and the knowledge cutoff is January 4, 2026.
Meta built it for agents that run locally: tool calls, multi-step plans and recovering from failed steps. Reasoning strength can be set to low, medium, high or xhigh. Full precision targets 64GB of graphics memory, while Meta's own quantised builds target 32GB and 24GB cards with average accuracy losses of 0.2% and 1% across 15 benchmarks. A bundled DFlash 'drafter' proposes 16 tokens at a time; Meta measured output rising from 74.9 to 233.4 tokens per second on an RTX 5090.
Everything, including the quantised weights and the drafter, is under Apache 2.0. Meta offers no API of its own. The weights are on Hugging Face, with GGUF and ExecuTorch builds, and OpenRouter lists hosted versions from DeepInfra and Together. On 15 October Meta also released a much larger open-weights model, Muse Spark Open, but under a custom licence with attribution and revenue terms.
Our review
Meta shipped Glimmer as a product, not a weights dump. The official quantised builds come with measured accuracy loss, the speculative-decoding drafter is included, and the speed table covers an RTX 5090 and two MacBook chips. A developer with a single 24GB card gets a tested configuration on day one, which few open releases bother to provide.
On Meta's own table it leads Gemma 4 31B and Qwen3.6-27B on agent tests: 75.5 on MCP Atlas against 54.2 and 62.5, and 51.2 on SWE-Bench Pro against 36.9 and 50.2. Qwen3.6-27B still won OSWorld-Verified, 75.6 to 65.9, and Terminal-Bench 2.1, 60.7 to 51.7. Then Qwen3.8-27B arrived and Alibaba's card puts it at 84.3 on OSWorld-Verified, nearly 20 points clear of Glimmer. All of these are vendor figures, run on each vendor's harness.
The security numbers deserve more attention than they get. On Meta's own privacy test Glimmer leaked in 26.4% of cases against Gemma's 12.1%, and on the AgentDojo prompt-injection test attacks succeeded 28.4% of the time. For a model sold as an always-on local agent with tool access, that argues for confirmation prompts on anything irreversible, which is what Meta's card recommends.
Price tilts the comparison further. OpenRouter lists Glimmer from $0.30 per million input tokens, while Qwen3.8-27B is listed at $0.04, and Qwen's 262K native context is double Glimmer's 131K. The clean Apache 2.0 licence from a US lab and the hardware packaging are real advantages, but they are now Glimmer's main ones.
Where it earns its place
A solo developer with an RTX 4090 or 5090 building a desktop automation agent
Run a tool-calling agent entirely offline, with no API bill
Load Meta's 17GB quantised build plus the drafter in llama.cpp; Meta measured 233.4 tokens a second on a 5090. The cost is the card; the risk is prompt injection, which succeeded 28.4% of the time on Meta's AgentDojo run.
A compliance-minded engineering lead at a US healthcare software company
Pick an open model whose licence and origin legal will sign off on
Apache 2.0 with no acceptable-use policy beyond the law, from a US lab, avoids the custom terms of Muse Spark Open and the questions some buyers raise about Chinese models. Patient data stays on hardware the company controls.
A startup running a few hundred internal agent sessions a day
Prototype on a hosted endpoint before buying GPUs
Call it through OpenRouter: 300M input and 60M output tokens a month is about $162 on DeepInfra. The same volume on Qwen3.8-27B at $0.04/$1.35 is about $93, so Glimmer's hosted endpoint makes sense mainly as a staging step before moving the same weights onto your own card.
Before you commit
- Setup
- Download the GGUF build from Hugging Face and run it in llama.cpp, Ollama or LM Studio; set reasoning strength in the system prompt with 'Reasoning strength: high'.
- Integrations
- Transformers, vLLM and llama.cpp; ExecuTorch builds for Apple silicon; works with OpenClaw and Hermes Agent scaffolds; OpenRouter for hosted calls.
- Data and privacy
- Apache 2.0 permits commercial use with no revenue cap. Run locally, nothing leaves the machine; on OpenRouter the hosting provider's terms apply. Meta's training data includes information from its own products and services.
- Cost at scale
- Local: one 24GB card for the 17GB build, 32GB for the near-lossless build, 64GB for full precision. Hosted: $0.30/$1.20 per 1M tokens on DeepInfra, $0.35/$1.50 on Together, so 300M in and 60M out a month is about $162.
- Limits
- 131,072-token context; text and image input only, video treated as frames; vendor-reported benchmarks only; 26.4% violation rate on Meta's privacy test and 28.4% attack success on AgentDojo.
What's good
- Apache 2.0, including quantised weights and drafter
- Official builds for 24GB and 32GB cards with 1% and 0.2% measured accuracy loss
- Bundled drafter lifts an RTX 5090 from 74.9 to 233.4 tokens a second, per Meta
- 75.5 on MCP Atlas, ahead of Gemma 4 31B and Qwen3.6-27B on Meta's table
What's not
- Qwen3.8-27B scores 84.3 to its 65.9 on OSWorld-Verified, per Alibaba
- 131K context, half Qwen3.8-27B's native 262K; no audio, video only as frames
- Leaks more than Gemma on Meta's own privacy test (26.4 against 12.1)
- No training data or code released
Best for
Developers building local, always-on agents on one consumer GPU who want a permissive licence from a US lab.
Alternatives
Muse Glimmer: common questions
- Is Muse Glimmer worth it?
- Glimmer is the open model most carefully fitted to the graphics card people own, and Qwen3.8-27B beat it on the benchmarks within weeks. We rate it 4.1 out of 5.
- How much does Muse Glimmer cost?
- Free weights under Apache 2.0; no Meta API. Hosted on OpenRouter from $0.30 per 1M input and $1.20 per 1M output tokens (DeepInfra).
- What are the downsides of Muse Glimmer?
- Qwen3.8-27B scores 84.3 to its 65.9 on OSWorld-Verified, per Alibaba; 131K context, half Qwen3.8-27B's native 262K; no audio, video only as frames; leaks more than Gemma on Meta's own privacy test (26.4 against 12.1); no training data or code released.
- Who is Muse Glimmer best for?
- Developers building local, always-on agents on one consumer GPU who want a permissive licence from a US lab.
- How much does Muse Glimmer cost for a team or in production?
- Local: one 24GB card for the 17GB build, 32GB for the near-lossless build, 64GB for full precision. Hosted: $0.30/$1.20 per 1M tokens on DeepInfra, $0.35/$1.50 on Together, so 300M in and 60M out a month is about $162.
- Is Muse Glimmer safe for company data?
- Apache 2.0 permits commercial use with no revenue cap. Run locally, nothing leaves the machine; on OpenRouter the hosting provider's terms apply. Meta's training data includes information from its own products and services.
- What are the best Muse Glimmer alternatives?
- The closest alternatives we have reviewed are Qwen3.8-27B, Gemma 4, gpt-oss-20b and Muse Spark 1.3.
Answers checked Oct 22, 2026.
Make Muse Glimmer? Show this rating on your site
