Meta drops a 30B-param local agent model that beats Qwen and Gemma on key benchmarks

Meta’s Superintelligence Labs just open-sourced Muse Glimmer, a 30-billion-parameter model designed to run AI agents entirely on a consumer GPU. Released under Apache 2.0 on Hugging Face, the model targets developers who need on-device agents that can access private data — calendars, messages, files — without ever phoning home to a cloud server.

The benchmark table tells a compelling story. Muse Glimmer posted a 75.5 on MCP Atlas, a test of multi-turn agent scaffolding, while Google’s Gemma4-31B managed just 54.2 and Qwen3.6-27B hit 62.5. On DeepSearch QA, Meta’s model scored 74.6 against Qwen’s 71.1. The gaps narrow considerably in coding tasks — Muse Glimmer led SWE-Bench Pro at 51.2, but Qwen3.6-27B edged it out on SWE-Bench Verified with 77.2 versus 76.0. These aren’t crushing victories. They’re competitive jostling in a field where model weights are now table stakes.

What Meta is really selling here is an architecture, not just a model. Muse Glimmer includes a perception encoder for interleaved text and images, which matters for agents that need to read screenshots, charts, or scanned documents as part of a workflow. The model supports OpenClaw and custom orchestration scaffolds, and Meta’s documentation describes retry training for failed tool calls — a practical feature for anyone who’s watched an agent loop endlessly on a broken command. The multimodal benchmarks are tight, with Qwen leading ScreenSpot Pro and OmniDocBench by razor-thin margins.

Safety numbers are worth a close read. On the CI Memories test, Muse Glimmer’s violation rate sits at 26.4 — notably higher than Gemma4-31B’s 12.1 but far lower than Qwen3.6-27B’s 53.4. The Siren AgentDojo attack success rate follows a similar pattern: 28.4 for Meta’s model versus 40.3 for Qwen. This isn’t a model you’d trust with a corporate credit card without guardrails, but it suggests Meta put real work into reducing prompt-injection vulnerability rather than treating safety as an afterthought. For teams that have been waiting to run capable agents without sending every keystroke through someone else’s API, this release lowers the barrier considerably — assuming they’re ready to build the scaffolding that makes a model like this actually useful.