Muse Glimmer 30B is a new 30-billion-parameter causal language model from the Meta Superintelligence Lab, distilled directly from Muse Spark. Released under the permissive Apache 2.0 license, it is purpose-built to execute complex, autonomous agentic workflows entirely on consumer hardware without relying on cloud infrastructure.
You can read more about the model on AIModels.fyi here!
Key Highlights
-
Built for Autonomous Agents: Engineered for end-to-end task completion, multi-step reasoning, and reliable tool use. It actively diagnoses errors and recovers from failed tool calls instead of halting.
-
Multimodal Perception: Features a dedicated ~1.8B parameter ViT-G/14 perception encoder, allowing the model to natively interleave text and images (such as interpreting charts, screenshots, and documents) within conversational workflows.
-
Optimized for Consumer Hardware: Designed to run locally on 24GB or 32GB VRAM setups (like a Mac or a single consumer GPU). Meta uses 4-bit quantization to shrink the language model to under 20GB with minimal degradation.
-
Fast Speculative Decoding: Ships with a lightweight “DFlash” block-diffusion drafter model that proposes 16 tokens at once, delivering speedups of up to 3.1x on an RTX 5090 and making it fast enough for real-time interaction.
-
Controllable Reasoning: Developers can explicitly set the model’s reasoning strength (
low,medium,high,xhigh) via the system prompt to balance speed and response quality for complex coding and problem-solving tasks. -
Strong Benchmark Performance: Competes fiercely in its size class, outperforming comparable models (like Gemma4-31B and Qwen3.6-27B) on key agentic and coding evaluations, including SWE-Bench Verified (76.0%) and MCP Atlas (75.5%).
License & Availability:
Muse Glimmer 30B is available under the Apache 2.0 license for commercial and research use. The Hugging Face release includes full-precision BF16 weights, 4-bit quantized variants, the DFlash drafter head, and the frozen perception encoder.
