Muse Glimmer lands as a 30B local agent

Meta Superintelligence Labs has shipped its first open model. Muse Glimmer is a 30-billion-parameter multimodal release, licensed under Apache 2.0, built from the ground up for agents that live on your laptop rather than in someone else’s data centre. Weights are live on Hugging Face today; optimised runtimes for llama.cpp, MLX and ExecuTorch are rolling out across Ollama, LM Studio, Unsloth and vLLM over the next few days.

This is the model Meta’s new lab has been pointing at since its launch — a compact, distilled-from-Muse-Spark build that the team explicitly tuned for the things you can’t do well over a slow connection: long-horizon coding agents, back-to-back tool calls, and the kind of personal-assistant pattern where your AI keeps a coherent thread going across the working day.

1.5×–1.8×faster generation on Apple Silicon via the new DFlash drafter in Ollama’s MLX engine.

What Meta released

Muse Glimmer arrives as a 30B multimodal model with a 1.8-billion-parameter perception encoder bolted on for image understanding. The combination — language model, encoder, drafter — has been sized so it all lives comfortably inside a single consumer GPU at the right quantisation. Meta’s research blog spells out the design intent: agent workloads that have to work whether the network is up or not, on hardware a small team can already afford.

The training pipeline reads like a deliberate play for that audience. Pre-training distilled logits from Meta’s larger Muse Spark teacher; mid-training added longer-context, agent-heavy data with richer reasoning traces; post-training mixed supervised fine-tuning with on-policy distillation and reinforcement learning across general, reasoning, coding and agentic domains. Meta’s own framing: capability balanced against the memory and compute constraints of local hardware.

Where it lands first

Ollama shipped its integration the same morning. On Apple Silicon, the new MLX engine adds two things alongside the model itself: native image input, and DFlash — a speculative-decoding drafter that proposes blocks of tokens for the main model to verify in parallel. Meta’s blog describes DFlash as a small companion network; Ollama measures it at 1.5×–1.8× faster generation in the MLX path. NVIDIA and AMD optimisations are flagged for the coming days.

The Ollama blog is blunt about the target workload: coding agents. They publish one-line launches for Claude Code, Pi, OpenClaw and Hermes, plus Codex, OpenCode and GitHub Copilot via the same ollama launch interface. Pi in particular is framed as the lighter-weight coding option.

How it compares at 30B

The numbers on the model card are what make this interesting against the existing local-agent field. SWE-Bench Verified sits at 76 for Muse Glimmer (ranked #14 overall, #2 in the under-32B size class on Hugging Face’s leaderboard); SWE-Bench Pro at 51.2. GPQA Diamond reads 83.5; MathArena AIME 2026 hits 94.7. ScreenSpot-Pro overall is 75.4 — #1 in its size bracket for screen-grounded tasks. Against the directly comparable 30-ish-billion open weights, Gemma 4-31B and Qwen 3.6-27B, Meta’s post positions Glimmer as the one to beat for always-on agent work.

The agentic-specific evaluations matter more for the use case. The model card lists deep end-to-end scores on DeepSearch QA, MCP-Atlas, τ-Bench and SWE-Bench, plus tool-use, multi-step reasoning and failure-recovery training. Meta also calls out scaffold compatibility with OpenClaw and other agentic orchestration patterns, controllable reasoning strength (low, medium, high, xhigh), and training data covering more than 100 languages.

What to try this afternoon

The pitch from both Meta and Ollama is the same: a single command, then a single agent command. If you’re on a Mac with Apple Silicon, the fastest path is:

• Pull it: grab the latest Ollama release, then run ollama run muse-glimmer:30b-mlx in your terminal.
• Wire up a coding agent: run ollama launch claude —model muse-glimmer:30b-mlx — or swap claude for pi for the lighter option.
• Try a long-running assistant: ollama launch openclaw —model muse-glimmer:30b-mlx or ollama launch hermes —model muse-glimmer:30b-mlx.
• On Nvidia: wait a few days for the CUDA build before benchmarking against Qwen 3.6 27B on the same hardware.

Three limits worth flagging. First, image input is MLX-only today — Nvidia and AMD support “in the coming days”. Second, the model’s coding-agent chops look strong on the public benchmarks, but real workflow speed depends as much on the harness around the model as on the weights; if you’ve been frustrated by slow local agents before, the llama.cpp fix story and the cache-not-raw-speed pattern both still apply. Third, this is the first open release from Meta Superintelligence Labs; the surrounding toolchain (fine-tuning recipes, scaffold docs, partner optimisations) is thinner than the more mature open-weight incumbents, and you’ll be closer to the bleeding edge than to a polished default.

The bigger picture is that the local-agent class just got denser. A 30B model with native images, a working drafter, and one-line launches for the agents people actually use is a different kind of release than another chat-tuned checkpoint. If you’ve been holding off on a local coding agent because the last few generations felt slower than they needed to be — Glimmer, with DFlash, is the test worth running on your own laptop this afternoon.