A 30-billion-parameter model at full precision needs over 55 GB of memory - more than any consumer GPU offers. Meta quantizes Muse Glimmer to roughly 4-bit, shrinking it to under 20 GB. That one engineering choice is what makes today's release worth paying attention to. Muse Glimmer is designed to run agentic tasks on a Mac or PC with a single graphics card, aiming at demand for AI systems that run directly on people's devices.
It shipped this morning. The weights are on Hugging Face. Here is what it actually does, where it wins, and where the benchmark table hides a real problem.
What "local agent" means here versus a local chatbot
Most open-weight models optimized for consumer hardware are, functionally, fast chatbots. Most LLMs are optimized for chat, prioritizing single-turn interactions and fast time to first token - but agentic workloads demand a different approach. Muse Glimmer is trained end-to-end around the agent loop, not tuned afterward.
Muse Glimmer is a 30-billion-parameter causal language model with a dedicated perception encoder, distilled from Muse Spark and purpose-built for autonomous agentic tasks on consumer hardware. The model integrates multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery into a single model that runs locally without requiring cloud infrastructure or network access.
The failure recovery piece is more important than it sounds. When a tool call fails or returns an unexpected result, the model diagnoses the error and retries rather than halt. Most fine-tuned local models just stop. That difference is the gap between a demo and something that can actually sit inside a CI pipeline overnight.
The release targets an operational constraint facing AI teams: cloud-hosted models need network access and central infrastructure. Meta pitches Muse Glimmer for workloads that require an on-device model, including personal agents with access to schedules, messages, files, and other private context.
The benchmark numbers - and what they don't say
Meta ran a three-way comparison against Gemma4-31B and Qwen3.6-27B across agentic tasks and general benchmarks. Glimmer leads in the categories that matter most for multi-step coordination:
| Benchmark | Muse Glimmer | Qwen3.6-27B | Gemma4-31B |
|---|---|---|---|
| MCP Atlas | 75.5 | 62.5 | 54.2 |
| DeepSearch QA | 74.6 | 71.1 | 61.7 |
| SWE-Bench Pro | 51.2 | 50.2 | 36.9 |
| WildClawBench | 47.6 | 43.2 | 37.6 |
| GAIA2 | 43.3 | 40.0 | 36.4 |
| OSWorld-Verified | 65.9 | 75.6 | - |
| TerminalBench 2.1 | 51.7 | 60.7 | - |
| SWE-Bench Verified | 76.0 | 77.2 | - |
The pattern is consistent: Muse Glimmer wins on agentic orchestration and reasoning. It trails on computer-use and terminal work.
That split matters for how you route work. If your agent needs to orchestrate tool calls, parse documents, and hand off results to a human in Slack - Glimmer fits. If it needs to drive a terminal session or control a desktop, Qwen3.6-27B stays ahead on the benchmarks that simulate that.
Optimized to run across NVIDIA edge, desktop, and workstation AI platforms, Muse Glimmer delivers 20K tokens/sec on a single GPU
- fast enough for a real-time agent loop rather than a batch job you kick off before lunch.
The safety number nobody is foregrounding
Giving a local model access to tools creates a different security problem from deploying a local chatbot. Meta's own safety table shows the tension.
On Siren AgentDojo, a prompt-injection test, Muse Glimmer shows a 28.4% attack-success rate versus 25.6% for Gemma4-31B and 40.3% for Qwen3.6-27B - while posting the highest utility score of the three at 94.2.
On CI Memories, a privacy benchmark where lower violation rates are better, Glimmer records 26.4 against Gemma's 12.1 and Qwen's 53.4.
So Glimmer is the most capable of the three, the most exploitable of the two best safety performers, and somewhere in the middle on privacy leakage. That is not a disqualifying combination - but it means the model needs guardrails that a pure chatbot deployment does not. Regulated enterprises get an air-gappable agent, but Meta advises adding system-level guardrails rather than shipping the model as a bare endpoint.
For a Slack-connected agent with access to calendars, project docs, or HR data, a 28.4% prompt-injection success rate against an adversarial test suite is a number worth knowing before you deploy.
How to actually deploy Muse Glimmer
The Hugging Face collection carries BF16 weights, GGUF k-quants, ExecuTorch builds, and the DFlash drafter. That covers three distinct hardware paths:
NVIDIA GPU (24 GB+): GGUF via llama.cpp, or NIM containers. Supported hardware includes GeForce RTX 5090, DGX Spark, DGX Station, and Jetson, with deployment via NVIDIA NIM containers, SGLang, and vLLM.
Apple Silicon: ExecuTorch builds target M4 and M5 Max Macs specifically.
On-prem server: BF16 weights with vLLM if you have the VRAM budget to skip quantization.
Muse Glimmer works across OpenClaw, Hermes Agent, and other agentic orchestration patterns
- so if your team already runs an OpenClaw agent in Slack, the scaffold swap is a config change, not a rewrite.
A teammate like Beagle, connecting the model's outputs back into a Slack channel, would handle the approval layer: the model drafts a response or action, a human approves, and the result posts with a source log. That chain keeps a person in the loop even when the model is running entirely on-prem.
Beagle in action#engineering-agents, 10:02amMeta Muse Glimmer local agent: common questions
What hardware does Muse Glimmer require?
Muse Glimmer needs a 24 GB or 32 GB consumer GPU, or an Apple M4/M5 Max Mac. At 4-bit quantization the model fits under 20 GB, leaving room for the KV cache and speculative-decoding drafter. NVIDIA RTX 5090, DGX Spark, and Jetson are all confirmed deployment targets.
How does Muse Glimmer compare to Llama models for agents?
Muse Glimmer is distilled from Muse Spark, not the Llama family. It is purpose-trained on agentic tasks end-to-end rather than fine-tuned afterward, which separates it architecturally from the Llama lineage. Its MCP Atlas and DeepSearch QA scores lead Gemma4-31B and Qwen3.6-27B in its size class, though Qwen leads on terminal and computer-use tasks.
Is Muse Glimmer safe enough to give access to internal files?
With caveats. Meta's own safety evaluation shows a 28.4% prompt-injection attack-success rate on Siren AgentDojo - better than Qwen's 40.3% but worse than Gemma's 25.6%. Meta explicitly recommends adding system-level guardrails before connecting the model to private data. Treat the safety score as a baseline, not a ceiling.
What is the Muse Glimmer license?
Apache 2.0, with weights on Hugging Face. This allows commercial use and modification. The license covers BF16 weights, GGUF quantizations, and ExecuTorch builds.
Can Muse Glimmer work with MCP servers?
Yes. The model scored 75.5 on MCP Atlas - the highest in its size class - and is explicitly trained on MCP-style scaffolds. It works with the July 28, 2026 stateless MCP spec, meaning you can run it behind a standard load balancer without sticky sessions.