Meta Releases Muse Glimmer 30B Open-Weight Model, Runs on a Single Consumer GPU

Meta has released the Muse Glimmer 30B-parameter multimodal model with weights open-sourced under the Apache 2.0 license. Distilled from Muse Spark, the model is optimized for always-on local agent workflows and can run on a single consumer GPU or Mac without network calls.

Facts

Muse Glimmer has approximately 30B total parameters, including a vision tower. The model uses a dense causal Transformer architecture with grouped-query attention using 32 query heads and 2 KV heads, with an attention pattern of [Local, Local, Local, Global] and a sliding window of 2048. The vision encoder is an approximately 1.8B ViT-G/14 supporting up to 4,096 visual tokens per image. Context length is 131,072+, vocabulary is 202,048, and the knowledge cutoff is January 4, 2026. It accepts text and image inputs and generates text output.

Training was conducted in three stages. Pre-training used logit distillation from Muse Spark outputs. Mid-training incorporated longer contexts and agent-dense data with richer reasoning traces. Post-training combined supervised fine-tuning, policy distillation, and reinforcement learning, covering general, reasoning, coding, and agent domains.

Technical Breakdown

At full precision, the model requires over 55GB of memory. Meta compressed the weights to approximately 4-bit precision, bringing the language model portion below 20GB, with the remaining space allocated to KV cache, perception encoder, and drafter. Two quantized versions have been released: K-Quant-Dynamic for 32GB VRAM with an average degradation of 0.2%, and K-Quant-17GB for 24GB VRAM with an average degradation of 1.0%. Block-level speculative decoding further improves speed, enabling the model to be embedded in real agent loops.

The model is available on Hugging Face with BF16 weights, GGUF k-quants, ExecuTorch builds, and a DFlash drafter. Optimized integrations for llama.cpp, MLX, and ExecuTorch are coming soon.

Industry Impact

For developers, the open weights and local run capability lower the barrier to entry. Solo developers and startups can deploy on a single 24GB GPU or M4/M5 Max Mac without paying per-token fees. Users have already tested its tool calling and multi-step reasoning in local multi-agent architectures.

For enterprise users, the model supports offline operation and data residency, making it suitable for scenarios such as healthcare, legal, financial services, defense, public sector, manufacturing, and field services. Regulated enterprises can implement air-gapped agent deployments. Meta recommends adding system-level guardrails rather than releasing the model as a bare endpoint.

For the competitive landscape, the open-weight license is more permissive than previous Llama licenses, offering a new option for local agent deployment. Simon Willison tested the 18.16GB quantized version on his blog for coding agents and image description, confirming it runs on machines with 32GB+ RAM alongside other applications.

Comparison and Precedents

The model achieves strong success rates on full-task benchmarks including DeepSearch QA, MCP-Atlas, τ-Bench, and SWE-Bench, measuring its ability to work within scaffolding, write and debug code, and handle multi-turn requests. For tool use, it handles a wide range of function calls, precisely invoking tools in extended workflows.

Strategic Assessment

Based on available facts, Muse Glimmer is most likely to gain rapid adoption in local agent toolchains.

Developer selection advice: for offline, multi-agent, or data-sensitive scenarios, prioritize testing the compatibility of its quantized versions with existing toolchains. Enterprises should validate actual performance on long-context agent tasks in a controlled environment before deciding whether to replace cloud-based solutions.