Technology company Meta released Muse Glimmer, a 30-billion-parameter agentic AI model distributed under the permissive Apache 2.0 license. Developed by Meta Superintelligence Labs, the model is designed to operate as a fully capable autonomous agent—including planning, tool invocation, self-verification, and failure recovery—while remaining compact enough to run locally on consumer hardware with as little as 24 GB of video memory.
The weights are available immediately on Hugging Face, with integrations for popular inference engines and platforms such as Ollama, LM Studio, vLLM, SGLang, Together AI, Fireworks AI, and OpenRouter scheduled to follow in the coming days.
The release extends Meta’s tradition of open-sourcing foundational AI research, this time targeting the growing demand for local, always-on agent workflows that do not depend on cloud connectivity or external infrastructure.
The new AI model was built using a bespoke architecture and a novel distillation recipe intended to transfer agentic reasoning from a significantly larger teacher model, referred to as Muse Spark, into a more efficient form factor. The training pipeline comprised three phases: pre-training via logit distillation on the teacher’s outputs; mid-training on extended-context, agent-heavy data enriched with reasoning traces; and 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 before release.
Benchmark results indicate 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. 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.
To enable practical local deployment, Meta applied quantisation 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 post Meta Releases Muse Glimmer With Open Weights, Targeting On-Device AI Agents Via Novel Distillation Pipeline appeared first on Metaverse Post.
