Today, we're introducing Muse Glimmer, the adjacent exemplary from Meta Superintelligence Labs, and unfastened sourcing the exemplary weights nether a permissive Apache 2.0 license.
Muse Glimmer is simply a 30-billion-parameter exemplary optimized for always-on section supplier workflows. It’s mini capable to tally connected a Mac aliases PC pinch a azygous user GPU, enabling usage cases that scope from section agents and usability calling, to section coding, and LLM-as-a-judge evaluation. Muse Glimmer delivers beardown capacity connected cardinal agentic usage cases and benchmarks compared pinch starring models successful its size category.
Foundation models person achieved singular capabilities crossed reasoning, codification generation, and instrumentality usage — yet astir deployments still dangle connected unreality infrastructure and web access. Running models locally enables you to usage AI anywhere, anytime, pinch aliases without an net connection. This is progressively viable: the unfastened root organization has shown that smaller models, erstwhile trained effectively, tin attack frontier-level capacity connected targeted tasks. Muse Glimmer is optimized for these section usage cases.
Keeping pinch our agelong contented of sharing basal AI research, we're releasing Muse Glimmer unfastened weights coming connected Hugging Face, on pinch developer documentation to thief you commencement building and moving your ain agents. Muse Glimmer is built to activity pinch the devices developers already use. Optimized integrations connected llama.cpp, MLX, and ExecuTorch will onshore successful the coming days, truthful you tin spell from download to moving supplier successful minutes.
How We Trained Muse Glimmer
An supplier that manages your schedule, drafts your messages, organizes your files, and learns really you activity needs heavy entree to individual context. It besides needs respective capabilities moving successful concert: long-horizon execution, precise instrumentality calling, multimodal understanding, long-context memory, and instruction following.
We designed Muse Glimmer to equilibrium capacity against the representation and compute constraints of section hardware. This required a compact architecture, a caller distillation look that transfers agentic reasoning from a overmuch larger coach model, and conclusion optimizations — including quantization — to meet latency expectations. We achieved this successful the pursuing phases:
- Pre-Training. We trained Muse Glimmer connected Muse Spark's outputs utilizing logit distillation, leveraging a akin information operation arsenic the teacher.
- Mid-Training. We trained the exemplary connected longer-context, much agent-heavy information pinch richer reasoning traces, alongside integrated data.
- Post-Training. We mixed supervised fine-tuning pinch a operation of on-policy distillation and reinforcement learning crossed general, reasoning, coding, and agentic domains.
Muse Glimmer was evaluated nether the standards group retired successful Meta's Advanced AI Scaling Framework and assessed for open-weight merchandise crossed each applicable categories.
Built for Agents: What Muse Glimmer Can Do
Building effective agents requires cardinal capabilities moving together to execute the user’s goals. Muse Glimmer is trained and evaluated crossed each of the following:
- End-to-end Agentic Task Completion. Muse Glimmer achieves beardown occurrence rates connected full-task benchmarks including DeepSearch QA, MCP-Atlas, 𝛕-Bench and SWE-Bench, which measurement its expertise to activity wrong scaffolds, constitute and debug code, and resoluteness multi-turn requests from commencement to finish.
- Reliable Tool Use. The exemplary handles a wide scope of usability calls, invoking devices pinch precise schemas passim extended workflows.
- Multi-Step Reasoning. Muse Glimmer chains reasoning complete agelong horizons, sustaining coherent plans crossed complex, extended workflows.
- Failure Recovery. When a instrumentality telephone fails aliases returns an unexpected result, the exemplary is trained to diagnose the correction and retry alternatively than halt.
- Multimodal Input and Reasoning. Through a dedicated cognition encoder, the exemplary accepts interleaved matter and images. This enables agents to construe screenshots, charts, and documents alongside conversation.
- Scaffold Compatibility. Muse Glimmer useful crossed OpenClaw and different agentic orchestration patterns.
- Controllable Effort. Muse Glimmer supports different reasoning strengths to prime the correct equilibrium betwixt value and speed.
- Multilingual. Muse Glimmer is trained connected information from much than 100 languages.
Performance
We evaluated Muse Glimmer crossed a wide scope of benchmarks to measure the divers capabilities required for effective autonomous supplier behavior. Compared pinch Gemma4-31B and Qwen3.6-27B, Muse Glimmer performs powerfully for its size people connected respective wide utilized LLM benchmarks.
For much item astir our evaluations, spot our report.
Optimized for Local Deployments
A section supplier is genuinely useful if it's accelerated capable to consciousness responsive. An supplier that takes minutes to reply aliases scheme its adjacent measurement breaks the travel of existent work. We applied 2 optimizations to make Muse Glimmer tally astatine applicable speeds connected user hardware without sacrificing quality.
Fitting the Model connected Your Device.
At afloat precision, a 30-billion parameter exemplary would require complete 55 GB of representation — acold much than immoderate user GPU offers. We usage quantization techniques to compress the model's weights to astir 4-bit precision, shrinking the connection exemplary to nether 20 GB. This leaves capable headroom for the model's moving representation (its "KV cache"), the cognition encoder for image understanding, and the speculative decoding drafter to tally simultaneously wrong a 24 GB aliases 32 GB envelope. We validated that this compression introduces minimal to nary degradation connected agentic tasks.
Faster Generation Through Speculative Decoding.
Language models usually make matter 1 token astatine a time, which tin consciousness slow during agelong reasoning chains aliases multi-step instrumentality calls. Muse Glimmer ships pinch a lightweight "drafter" exemplary based connected DFlash — a mini companion web that proposes full blocks of tokens astatine once. The main exemplary past verifies these proposals successful parallel, accepting correct tokens and correcting incorrect ones. This method lets Muse Glimmer make matter importantly faster than modular token-by-token procreation while producing identical output quality. We supply quantized drafter versions to incur a smaller representation overhead successful the release.
The Result:
We measurement the velocity of our K-Quant-17GB exemplary alongside the quantized DFlash drafter connected MacBook M4-Max, M5-Max and connected a RTX-5090. The exemplary is accelerated capable for fluid speech and real-time supplier interaction, each moving wholly connected your device.
Get Started With Muse Glimmer Today
Muse Glimmer is disposable now, and you tin download the weights connected Hugging Face. In the coming days, tally it locally done partners for illustration Ollama, LM Studio, and Unsloth, deploy it pinch separator frameworks including llama.cpp, ExecuTorch, and MLX, service it astatine standard pinch vLLM and SGLang, aliases get started quickly done partners for illustration Together AI, Fireworks AI, and OpenRouter. You tin moreover customize it for your use-case by leveraging PyTorch’s TorchTitan training characteristic to customize the exemplary further.
We're besides moving pinch our partners including AMD, Arm, Dell, Intel, and NVIDIA to optimize capacity crossed devices. In addition, we’re releasing archiving truthful developers person the resources they request to get started and build responsibly pinch Muse Glimmer. This includes guidance connected mounting up civilization scaffolds, truthful it's moreover easier to commencement building and deploying individual agents connected time one. You tin study much and find resources to build connected Meta's AI Developer Center.
This activity builds connected Meta's agelong way grounds of unfastened AI research, extending it into agentic AI and giving developers entree to section agentic capabilities. As always, we invited feedback from the organization and can’t hold to spot what developers build pinch this unfastened weights model.
Download the Model connected HuggingFace Developer Documentation
English (US) · 
Indonesian (ID) ·