Meta is releasing Muse Glimmer under an Apache 2.0 licence for local AI agents that can run on a consumer GPU.
The company’s Superintelligence Labs has released the 30-billion-parameter model’s weights on Hugging Face. Meta says developers can use it for local coding, function calling, local agents, and LLM-as-a-judge evaluation.
The release targets an operational constraint facing AI teams: cloud-hosted models need network access and central infrastructure. Meta instead pitches Muse Glimmer for workloads that require an on-device model, including personal agents with access to schedules, messages, files, and other private context.
Meta Muse Glimmer leads several agent task benchmarks
Meta’s benchmark tests put Muse Glimmer ahead of Gemma4-31B and Qwen3.6-27B on five of eight general-agentic benchmarks. The model scored 75.5 on MCP Atlas. Gemma4-31B reached 54.2, and Qwen3.6-27B recorded 62.5.
DeepSearch QA follows a similar pattern. Meta reports a score of 74.6 for Muse Glimmer, against 61.7 for Gemma4-31B and 71.1 for Qwen3.6-27B. The supplied announcement identifies both benchmarks as tests of an agent’s ability to work within scaffolds and complete multi-turn requests.
The model scored 23.5 on τ²-Banking. Gemma4-31B recorded 15.1. Qwen3.6-27B reached 16.7.
Muse Glimmer also posted 47.6 on WildClawBench, ahead of Gemma4-31B’s 37.6 and Qwen3.6-27B’s 43.2. Its GAIA2 result reached 43.3, compared with 36.4 and 40.0 respectively.
Other agent scores favour Qwen3.6-27B. Meta’s table gives that model 1,141 on GDPval-AA, against Muse Glimmer’s 953 and Gemma4-31B’s 811. Qwen3.6-27B also led SkillsBench with Skills at 46.6, where Muse Glimmer recorded 44.3.
OSWorld-Verified produced the largest gap in this group. Meta reports 75.6 for Qwen3.6-27B. Muse Glimmer reached 65.9, and Gemma4-31B scored 58.5.
These tests measure constrained tasks. They do not demonstrate how a local agent will behave after an organisation connects it to its own files, calendars, messaging systems, or internal tools.
Coding results split between Muse Glimmer and Qwen
Muse Glimmer’s coding results show a narrower comparison. The model led SWE-Bench Pro with a score of 51.2. Meta reports 36.9 for Gemma4-31B and 50.2 for Qwen3.6-27B.
SciCode produced a close result. Muse Glimmer scored 43.6, marginally above Gemma4-31B at 43.4. Qwen3.6-27B recorded 39.8.
Qwen3.6-27B led two other coding evaluations. It scored 77.2 on SWE-Bench Verified, compared with Muse Glimmer’s 76.0. TerminalBench 2.1 gave Qwen3.6-27B a score of 60.7; Muse Glimmer reached 51.7, and Gemma4-31B posted 43.4.
A local coding agent does more than produce code. It needs a scaffold that decides which repositories, terminals, test environments, and commands the model may access. Meta says Muse Glimmer supports OpenClaw and other agent-orchestration patterns, with custom scaffolds covered in its developer documentation.
An organisation evaluating the model for software work should define the commands and repositories available to the agent before measuring task success. The supplied material describes retry training for failed tool calls. That behaviour requires controls over repeat attempts, especially where a tool can alter source code or invoke an external system.
Multimodal scores favour Qwen in most tests
Muse Glimmer accepts interleaved text and images through a dedicated perception encoder. Meta says this design lets agents interpret screenshots, charts, and documents as part of a conversation.
The benchmark chart puts Muse Glimmer ahead on Charxiv Reasoning. Its score reached 78.8, against 77.7 for Gemma4-31B and 78.4 for Qwen3.6-27B.
Qwen3.6-27B led ScreenSpot Pro with 76.1. Muse Glimmer recorded 75.4, and Gemma4-31B scored 75.9. The same model led OmniDocBench v1.5 at 77.8, compared with Muse Glimmer’s 75.8 and Gemma4-31B’s 72.5.
MMMU Pro produced smaller differences. Meta lists Muse Glimmer at 74. Qwen3.6-27B reached 75, and Gemma4-31B posted 73.
These results matter for teams considering agents that act on visual interfaces. A screenshot-reading model can interpret what it sees, yet local testing must still cover permissions, display layouts, document formats, and errors returned by connected tools.
Safety figures show lower reported attack success than Qwen
Meta also reports two safety-related evaluations: CI Memories and Siren AgentDojo. The chart uses different measures for each test.
On CI Memories, Meta lists a violation rate of 26.4 for Muse Glimmer and a coverage score of 64.8. Gemma4-31B recorded a violation rate of 12.1 with coverage of 53.0. Qwen3.6-27B posted a violation rate of 53.4 and coverage of 66.9.
The Siren AgentDojo result uses attack success rate and utility. Meta gives Muse Glimmer an attack success rate of 28.4 and a utility score of 94.2. Gemma4-31B scored 25.6 on attack success rate, with utility at 90.8. Qwen3.6-27B recorded 40.3 and 92.7.
General reasoning results add context to agent claims
Muse Glimmer led four of six general-capabilities-and-reasoning tests in Meta’s comparison. It scored 77.0 on IFBench. Gemma4-31B recorded 76.0, and Qwen3.6-27B reached 70.8.
The AIME 2026 score was 94.7 for Muse Glimmer. Meta reports 89.2 for Gemma4-31B and 94.1 for Qwen3.6-27B. On AA-LCR, Muse Glimmer reached 80.0, ahead of 68.3 and 73.3.
The model also led Beam 128K at 65.1. Qwen3.6-27B scored 63.0. Gemma4-31B recorded 58.2.
Gemma4-31B led GPQA Diamond with 85.7. Muse Glimmer scored 83.5, followed by Qwen3.6-27B at 84.2. Gemma4-31B also took the top score on Humanity’s Last Exam, Text No Tools, at 23.6; Muse Glimmer reached 22.0.
One model does not lead every test. Meta’s results instead show Muse Glimmer competing closely with two similarly sized models across a mixed set of agent, coding, visual, safety, and reasoning evaluations.
Memory limits shape the local deployment design
Meta says a full-precision 30-billion-parameter model would require more than 55 GB of memory. Muse Glimmer instead uses approximately 4-bit weight quantisation, reducing the language model to under 20 GB.
That allocation leaves memory for a KV cache. The model also needs room for its perception encoder and a speculative-decoding drafter. Meta targets a 24 GB or 32 GB memory envelope for these components.
The company says the DFlash-based drafter proposes blocks of tokens for the main model to verify in parallel. Meta says this speeds generation compared with standard token-by-token output and retains identical output quality. The supplied post does not include token-per-second figures, prompt sizes, power data, or concurrency results.
Meta tested its K-Quant-17GB version with the quantised DFlash drafter on MacBook M4-Max hardware, MacBook M5-Max hardware, and an RTX-5090. It describes the resulting experience as suitable for fluid conversation and real-time agent interaction.
The public weights are available through Hugging Face. Meta says integrations with llama.cpp, MLX, and ExecuTorch will arrive in the coming days.
See also: Alibaba tests new business model for Qwen open-source AI
Want to learn more about AI and big data from industry leaders? Check out AI & Big Data Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events including the Cyber Security & Cloud Expo. Click here for more information.
AI News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.
Read the full article here
