Meta releases Muse Glimmer, a 30-billion-parameter open model for local AI agents

Meta has released a 30-billion-parameter AI model under an open-source license, pitched for software agents that run on a home graphics card instead of a cloud server.

The model, called Glimmer. The weights are published on Hugging Face, and Meta's research blog frames the use case as local coding, function calling, and personal agents that can read a user's calendar, messages, and files without that data leaving the machine.

That on-device pitch is what separates the release from a generic open-model announcement. Most "agentic" AI today routes through a hosted API: prompts and tool calls leave the device, and the model has no access to local context unless the user pipes it in. A 30-billion-parameter model that fits on a consumer GPU lets a developer point the agent at a file system, a calendar, or a messaging client without the round trip. It also means the privacy story depends on what the developer wires into the model, not on the model itself.

The benchmark table, as reported by trade-press coverage of the release, is uneven. On agent-scaffold evaluations, tests that measure whether a model can call the right tool and chain the right steps, Muse Glimmer leads: MCP Atlas, a tool-use benchmark, at 75.5 versus 54.2 for Gemma4-31B and 62.5 for Qwen3.6-27B; DeepSearch QA, a multi-step web search and synthesis test, at 74.6; τ²-Banking, a domain-specific agent benchmark, at 23.5; WildClawBench, a long-horizon agent test, at 47.6; and GAIA2, a general assistant benchmark, at 43.3. On SWE-Bench Pro, a real-world software engineering test, Muse Glimmer posts 51.2 against Qwen3.6-27B's 50.2. On SciCode, a scientific code generation benchmark, it edges the field at 43.6.

Qwen3.6-27B wins the rest. It leads on OSWorld-Verified, a desktop operating-system control benchmark, at 75.6 versus Muse Glimmer's 65.9, on SWE-Bench Verified, a cleaner software-engineering set, at 77.2 versus 76.0, on GDPval-AA, an agentic eval tied to economically valuable tasks, at 1,141 versus 953, and on SkillsBench, a skills-assessment test, at 46.6 versus 44.3.

The split tracks a familiar pattern. Local models tend to win on tool-use and retrieval-style tasks where the test can be run on a single machine; hosted competitors still lead on long-horizon OS control and code-editing workflows, where latency and tool breadth matter more than local execution. None of the comparison model names (Gemma4-31B, Qwen3.6-27B) have been independently cross-checked against Meta's primary benchmark table, and the trade-press benchmark table is truncated, so the precise margins should be read as vendor-supplied until Meta publishes the full methodology.

Meta itself flags the same caveat. The release notes that benchmark scores do not demonstrate how a local agent behaves once an organization connects it to its own files, calendars, messaging systems, or internal tools. In other words, the table tells you the model can call a function or read a file in a sandboxed test. It does not tell you what happens when the agent is pointed at a real user's inbox.

The "consumer GPU" line deserves the same caution. A 30-billion-parameter model at standard precision needs tens of gigabytes of memory, well above the 8 to 24 GB that most consumer graphics cards ship with. Meta's release assumes some level of quantization, a technique that shrinks the model's weights to a lower-precision format at the cost of some accuracy, to fit a 30B model on a single consumer card. The exact memory floor will depend on which quantization the developer picks. Treat "runs on a home GPU" as a target, not a guarantee.

The license is the part that is straightforward. Apache 2.0 lets a developer use, modify, and redistribute the weights, including in commercial products. For a team building a local agent product, that removes one of the recurring friction points from the open-weights stack.

The cloud-agent build-out is the implicit target. OpenAI, Anthropic, and Google pitch hosted infrastructure as the natural home for agentic workloads: managed tools, persistent memory, and orchestration. Meta's position is that a slice of those workloads, the ones that touch personal context, treat the network round trip and the central infrastructure as a liability rather than a feature. A 30B open-weight model that fits on a home GPU does not settle that argument, but it makes the on-device side of it concrete for the first time at this scale.