Meta Releases Muse Glimmer: A 30B-Parameter Open-Weight Model Built for On-Device AI Agents
Key Takeaways
- •Muse Glimmer is a 30-billion-parameter agentic AI model released by Meta under the Apache 2.0 license, permitting commercial use, modification, and redistribution.
- •The model was trained through a three-phase pipeline that used logit distillation from a larger teacher model called Muse Spark, followed by agent-heavy mid-training and post-training with reinforcement learning.
- •Benchmark results show Muse Glimmer performing competitively against similarly sized models such as Gemma4-31B and Qwen3.6-27B on agent-specific workloads including multi-turn tool use, web-search QA, and software engineering tasks.
- •Meta applied approximately 4-bit quantization to compress the model to under 20 GB, enabling real-time on-device operation on consumer hardware including MacBook M4-Max, M5-Max, and RTX-5090 systems.
- •The model supports multimodal input through a dedicated perception encoder, operates across more than 100 languages, and is compatible with agentic orchestration frameworks such as OpenClaw.
Meta has released Muse Glimmer, a 30-billion-parameter agentic AI model distributed under the permissive Apache 2.0 license, which permits commercial use, modification, and redistribution with minimal restrictions. Developed by Meta Superintelligence Labs, the model is designed to function as a fully capable autonomous agent—handling planning, tool invocation, self-verification, and failure recovery—while remaining compact enough to run locally on consumer hardware requiring as little as 24 GB of video memory.
Model weights are available immediately on Hugging Face. Integrations with widely used inference engines and platforms—including Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI, and OpenRouter—are scheduled to follow in the coming days.
The release extends Meta's ongoing practice of open-sourcing foundational AI research, this time aimed at the increasing demand for local, always-on agent workflows that operate without cloud connectivity or external infrastructure. Running agents entirely on-device eliminates per-token API costs, removes network latency from agent reasoning loops, and keeps sensitive data from leaving the user's machine—a combination that matters for privacy-sensitive enterprise deployments, offline environments, and applications where sub-second tool-call latency is critical. The move also intensifies competition in the open-weight AI ecosystem, where Meta's Llama family, Google's Gemma line, Alibaba's Qwen series, and Mistral's models vie for developer adoption across both cloud and edge deployment scenarios.
Introducing Muse Glimmer, an open-weight 30B-parameter model optimized for local, always-on agent workflows. Muse Glimmer delivers strong performance on key agentic use cases and benchmarks compared with leading models in its size category, and is designed to run entirely on… pic.twitter.com/mI4z91GPnE
— AI at Meta (@AIatMeta) August 10, 2026
Architecture, Training, and Local Optimization
Muse Glimmer was built using a bespoke architecture and a novel distillation recipe designed to transfer agentic reasoning from a significantly larger teacher model—referred to as Muse Spark—into a more efficient form factor. Distillation has become a standard technique in the model-compression toolkit, enabling smaller models to approximate the output distributions and reasoning patterns of much larger systems. The training pipeline consisted of three phases:
- Pre-training via logit distillation on the teacher model's outputs.
- Mid-training on extended-context, agent-heavy data enriched with reasoning traces.
- Post-training combining supervised fine-tuning with on-policy distillation and reinforcement learning across general, coding, and agentic domains.
The model was evaluated under Meta's Advanced AI Scaling Framework prior to release.
Benchmark Performance and Capabilities
Benchmark results show competitive performance relative to similarly sized counterparts, including Gemma4-31B and Qwen3.6-27B, on tasks such as DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench. Notably, the evaluation suite emphasizes agent-specific workloads—multi-turn tool use, web-search-driven QA, and real-world software engineering tasks—rather than static knowledge benchmarks, reflecting the industry's broader shift toward measuring models on dynamic agentic capability rather than factual recall alone.
Beyond core reasoning, Muse Glimmer supports multimodal input through a dedicated perception encoder, multilingual operation across more than 100 languages, and compatibility with agentic orchestration patterns such as OpenClaw.
Local Deployment and Quantization
To enable practical local deployment, Meta applied quantization techniques that compress the model to approximately 4-bit precision, reducing its footprint to under 20 GB. This leaves sufficient memory for the KV cache, image encoder, and a lightweight speculative decoding drafter based on DFlash, which proposes token blocks in parallel to accelerate generation without altering output quality.
Meta validated the setup on MacBook M4-Max, M5-Max, and RTX-5090 hardware, reporting speeds suitable for fluid conversation and real-time agent interaction entirely on-device. The choice of validation hardware spans Apple Silicon and high-end NVIDIA consumer GPUs, reflecting the two dominant consumer inference stacks and underscoring Meta's intent to make the model accessible across the most common developer workstations rather than data-center hardware.
