Muse Glimmer
An open-weight on-device agent model released by Meta with 30 billion parameters, distilled from the Muse Spark large model. Designed for on-device agent workflows, it supports efficient execution on NVIDIA hardware via ExecuTorch.
Core Facts
Muse Spark can be used as a backend via OpenCode Go.
Running Muse Glimmer at full precision requires 55 GB of VRAM, which is compressed to under 20 GB via 4-bit quantization, allowing it to run on consumer-grade GPUs with 24 GB of VRAM.
Meta has consistently adopted an open-source tendency in the AI field, continuously building a developer ecosystem through open models from the Llama series to Muse Glimmer.
Timeline (last 90 days)
In the 16-concurrency, 256-token-output test, Gemma 4 E2B achieves an output speed of 1887 tokens/second, and the models faster than it include Qwen2.5-0.5B (4176, 2.21x), Qwen3-0.6B (3210, 1.70x), Qwen2.5-1.5B (2780, 1.47x), and Qwen3-1.7B (2049, 1.09x).
In a C++ wireless skateboarding game test, Muse Glimmer's attempt to fix a compilation error caused the entire system to crash completely
In actual testing under agentic tasks, the performance of this quantized version of Muse Glimmer drops by only 0.2%, and only 1% when compressed to the 24GB version
On an RTX 5090, Muse Glimmer runs at 75 tokens/second without speculative decoding, and reaches 233 tokens/second when paired with a speculative decoding model
Muse Glimmer has a context length of 131,072 (approximately 130,000 tokens) and comes with an extendable token marker
Muse Glimmer has multimodal capabilities, with a built-in vision encoder that can accept both text and image inputs simultaneously
Muse Glimmer adopts the Apache 2.0 license
Muse Glimmer can run on a single consumer-grade GPU with 24GB VRAM or a Mac equipped with sufficient unified memory
Meta released an open-weight model named Muse Glimmer
Muse Spark can be used as a backend via OpenCode Go.
26 more timeline events
All Facts (20)
Muse Spark can be used as a backend via OpenCode Go.
85%VerifiedRunning Muse Glimmer at full precision requires 55 GB of VRAM, which is compressed to under 20 GB via 4-bit quantization, allowing it to run on consumer-grade GPUs with 24 GB of VRAM.
80%VerifiedMeta has consistently adopted an open-source tendency in the AI field, continuously building a developer ecosystem through open models from the Llama series to Muse Glimmer.
65%VerifiedMuse Glimmer 30B has 29.6 billion parameters and is a dense model.
65%VerifiedMuse Glimmer supports multimodal input and has a context window of 128,000 tokens.
65%UnverifiedWhen memory scales to hundreds of MB or even GB, a pure file-based approach may still need to be combined with indexing mechanisms and vector retrieval technologies to balance simplicity and performance
60%UnverifiedMuse Glimmer is distilled from the Muse Spark model, with post-training, supervised fine-tuning (SFT), and reinforcement learning (RL) layered on top.
60%UnverifiedIn the 16-concurrency, 256-token-output test, Gemma 4 E2B achieves an output speed of 1887 tokens/second, and the models faster than it include Qwen2.5-0.5B (4176, 2.21x), Qwen3-0.6B (3210, 1.70x), Qwen2.5-1.5B (2780, 1.47x), and Qwen3-1.7B (2049, 1.09x).
50%UnverifiedIn actual testing under agentic tasks, the performance of this quantized version of Muse Glimmer drops by only 0.2%, and only 1% when compressed to the 24GB version
50%UnverifiedMuse Glimmer has multimodal capabilities, with a built-in vision encoder that can accept both text and image inputs simultaneously
50%UnverifiedMuse Glimmer adopts the Apache 2.0 license
50%UnverifiedMuse Glimmer can run on a single consumer-grade GPU with 24GB VRAM or a Mac equipped with sufficient unified memory
50%UnverifiedMuse Glimmer has a context length of 131,072 (approximately 130,000 tokens) and comes with an extendable token marker
50%UnverifiedOn an RTX 5090, Muse Glimmer runs at 75 tokens/second without speculative decoding, and reaches 233 tokens/second when paired with a speculative decoding model
50%UnverifiedvLLM v0.28.0 adds support for models including Muse Glimmer, Ling 3.0 Flash, Dots3 NOTE, and Interns2mobius.
50%UnverifiedRunning via llama.cpp on a MacBook Pro with an M4 Pro chip and 24GB unified memory, the text decoding speed is 10.13 tokens/second, with each complete task taking 4 to 7 minutes.
50%UnverifiedOn SWE Bench Pro, Glimmer narrowly beat Qwen with 51.2 vs. 50.2, but Qwen surpassed it on SWE Bench Verified (Qwen 77.2 vs. 76) and Terminal Bench (Qwen 60.7 vs. 51.7).
50%UnverifiedAt its size, Glimmer does not win on raw coding ability; its advantage lies in tool-calling capability.
50%UnverifiedMeta's officially stated hardware requirement for running Muse Glimmer is a 24GB graphics card.
50%UnverifiedUnsloth compressed Muse Glimmer through quantization to run within 14GB of system memory, with the agent still functioning normally.
50%