Meta is releasing the weights for Muse Glimmer, a 30-billion-parameter model designed to run AI agents on high-end consumer hardware. The Apache 2.0 release aims to make local coding, function calling, visual interpretation and agent orchestration possible without sending every request to a cloud service.
That proposition is particularly relevant for personal agents connected to schedules, messages and files. Local execution can reduce reliance on network access and centralized infrastructure, although it does not automatically make an agent private or safe. Permissions, tool access and the software surrounding the model still determine what information it can reach and what actions it can take.
Meta’s own evaluation presents Muse Glimmer as competitive with similarly sized models rather than universally dominant.
Agent benchmarks show strengths and uneven results
Meta reports that Muse Glimmer led Gemma4-31B and Qwen3.6-27B on five of eight general agent benchmarks in its comparison. It scored 75.5 on MCP Atlas, compared with 54.2 and 62.5, and reached 74.6 on DeepSearch QA, ahead of 61.7 and 71.1.
Results varied elsewhere. Qwen3.6-27B led GDPval-AA, SkillsBench and OSWorld-Verified. Muse Glimmer’s score of 65.9 on the last of those sat above Gemma4-31B’s 58.5 but below Qwen’s 75.6.
Coding tests were similarly mixed. Muse Glimmer narrowly led SWE-Bench Pro with 51.2 and SciCode with 43.6. Qwen led SWE-Bench Verified and TerminalBench 2.1. Meta also reported competitive visual and reasoning results, with no single model winning every test.
These are vendor-supplied benchmark numbers for constrained tasks. They do not establish how an agent will perform after being connected to a company’s repositories, terminals, calendars or internal services. Scaffolding and permissions can matter as much as the base model.
Quantisation makes the local proposition practical
A full-precision model of this size would require more than 55 GB of memory, according to Meta. Muse Glimmer uses roughly four-bit weight quantisation to reduce the language model below 20 GB, leaving capacity for its context cache, perception encoder and speculative-decoding component.
Meta targets a total envelope of 24 GB or 32 GB. It tested a 17 GB quantised version with a DFlash-based drafter on MacBook M4 Max and M5 Max systems and an RTX 5090. The company describes the experience as fluid enough for conversation and real-time agent interaction.
However, the announcement does not provide token-per-second measurements, prompt lengths, energy consumption or concurrency data. Buyers and developers will need their own tests before treating the hardware claim as a production capacity plan.
The model accepts interleaved images and text through a separate perception encoder. That allows an agent to reason over screenshots, charts and documents, extending the local use case beyond text-only assistance.
Local execution shifts rather than removes risk
Muse Glimmer supports OpenClaw and other orchestration approaches, with additional integrations for llama.cpp, MLX and ExecuTorch promised after release. An organization deploying it for coding still needs to decide which repositories, commands and test environments the agent may use.
Meta says its training includes retries after failed tool calls. That can improve task completion, but repeated attempts become dangerous when a tool can modify code or invoke an outside system. Retry limits, approval gates and isolated execution should be tested as part of the application rather than assumed to be properties of the model.
Meta’s safety evaluations also resist a simple winner narrative. On Siren AgentDojo, the company reports an attack success rate of 28.4 for Muse Glimmer, versus 25.6 for Gemma and 40.3 for Qwen, while utility scores followed a different ordering. Such figures show trade-offs, not proof of secure deployment.
Open weights and a smaller memory footprint give developers more control over where an agent runs. They also transfer more responsibility to the operator. Muse Glimmer lowers the hardware barrier to local agents; it does not lower the need for careful boundaries around them.
