Meta Superintelligence Labs open-sources 30B agent model Muse Glimmer, runs on consumer GPU
Meta released Muse Glimmer, a 30B dense model under Apache 2.0 license, on August 10. It runs on consumer GPUs with 24GB VRAM and is positioned as a local agent foundation.
On the evening of August 10, Meta Superintelligence Labs officially launched the Muse Glimmer model, releasing its weights under the Apache 2.0 license. It is the first open-weight model from Meta since Llama 4 and the first to use Apache 2.0. Unlike the previous Llama custom license, Apache 2.0 allows enterprises to deploy it directly and build derivative works integrated into their own products.
Muse Glimmer is a 30B dense model distilled from Meta's flagship Muse Spark, includes around 1.8B parameters of vision encoder, supports text and image input, with 128K context. The model is optimized for continuous local agent workflows, with emphasis on multi-step reasoning, tool calling, and failure recovery. It can self-diagnose tool failures and retry.
For local deployment, Meta compressed the weights to under 20GB using 4-bit quantization, enabling it to run on consumer GPUs with 24GB or 32GB VRAM. The model pairs with a lightweight DFlash draft model for speculative decoding to accelerate inference with minimal quality loss. Meta has worked with AMD, Arm, Dell, Intel, and NVIDIA for device-level inference optimization.
On benchmarks, Muse Glimmer scored 75.5 on MCP-Atlas, well above Qwen3.6-27B's 62.5; 74.6 on DeepSearch QA versus Qwen3.6's 71.1; and 94.7 on AIME 2026 math, slightly above Qwen3.6's 94.1. However, Qwen3.6 still leads on OSWorld-Verified (75.6) and TerminalBench 2.1 (60.7). Weights are live on Hugging Face, with Ollama, LM Studio, vLLM, SGLang, llama.cpp, MLX, and ExecuTorch support rolling out within the week. OpenRouter, Together, and Fireworks also serve the model. Meta Chief AI Officer Alexandr Wang previewed that Muse Spark 1.2 weights will open in the coming weeks.