Meta open-sources Muse Glimmer

Meta released Muse Glimmer on August 10, 2026. It is a 30-billion-parameter model, published on Hugging Face under the Apache 2.0 license, and Meta Superintelligence Labs says it is built for always-on local agent workflows: software that runs on your own machine, calls tools, recovers from its own failures, and keeps every prompt on the device.

It is the open counterpart to Muse Spark 1.1, the closed multimodal model Meta launched on July 9 alongside the Meta Model API. Spark's weights stayed private, a notable break for the company behind the open Llama models. Glimmer is where the open-weights habit resumes, at the small end of the line. Meta says it was trained on Spark's outputs through logit distillation, meaning the small model was taught to match the large model's answers, with later stages using longer agent-heavy traces, then supervised fine-tuning and reinforcement learning.

What it is built to do

Meta names local agents, function calling, local coding, and acting as a judge of other models' work. Glimmer accepts images through a dedicated encoder, so an agent can read screenshots and charts. Reasoning effort can be turned up or down depending on the task, and the model handles more than 100 languages, per Meta.

What hardware you need

At full precision a 30-billion-parameter model needs more than 55 GB of memory, more than any consumer GPU offers. Quantization, which stores each weight with fewer bits, brings the language model under 20 GB at about 4-bit precision. Meta tested a 17 GB compressed build on a MacBook M4 Max, an M5 Max, and an Nvidia RTX 5090, and says it fits inside a 24 GB or 32 GB card with room left for working memory and the image encoder.

Speed comes from speculative decoding: a tiny companion network Meta calls DFlash proposes a block of tokens, and the main model checks them in parallel, keeping the ones that match. Meta reports 3.1 times faster generation on the RTX 5090, 1.8 times on the M5 Max, and 1.5 times on the M4 Max, with output identical to ordinary generation. The drafter ships with the release. No tokens-per-second figures were published; Meta's claim is that it is fast enough for conversation and real-time agent work entirely on the device.

Can you get it today

The weights are up now. Meta is not serving Glimmer on its own API; the cloud Muse is still Spark 1.1, through the Meta AI app, meta.ai, or the Meta Model API. Integrations for llama.cpp, MLX, and ExecuTorch are promised in the coming days, with Ollama, LM Studio, and Unsloth to follow, and hosted access through Together AI, Fireworks AI, and OpenRouter. So a reader with the hardware can start from the Hugging Face files today. Everyone else waits for the runtimes or a host.

How good is it

Meta evaluated Glimmer against Gemma 4 31B and Qwen 3.6 27B, the two open models closest to its size, across agentic, coding, multimodal, safety, and reasoning tests. The named benchmarks are SWE-Bench, which tests fixing real bugs in real code, τ-Bench, which tests an agent completing tasks while talking to a user, and DeepSearch QA and MCP-Atlas, which test research and tool use. Meta says the model performs competitively for its size class, but it published no scores, so there is nothing for anyone to check yet. The context window also went unstated. Until independent tests arrive, the capability picture is Meta's word alone.