Meta introduced an open agent model called Muse Glimmer with 30 billion parameters on August 10, 2026, which runs entirely locally on a single consumer graphics card. The company distills the model from its larger cloud model Muse Spark and makes it available for free under the Apache 2.0 license on Hugging Face. Unlike previous agent models, Muse Glimmer no longer needs a cloud connection.
Four-bit quantization cuts memory needs to one third
Muse Glimmer combines a 28-billion-parameter text decoder with a two-billion-parameter image encoder in the ViT style, which processes text, images, and short video sequences together. A changing mix of local attention windows and full attention, plus a method called “gated” key-value caching, reduces the memory needed for context by a factor of sixteen, according to Meta’s own research blog. With additional four-bit quantization, total memory demand drops from 55 to 18 to 20 gigabytes, letting the model, cache, and image encoder fit on a single graphics card with 24 or 32 gigabytes.
An extra component called DFlash speeds up text generation through speculative decoding by a factor of 1.5 to 3.1, reaching up to 20,000 tokens per second on one graphics card. At launch, the Transformers library, llama.cpp, and vLLM support the model directly, as do Hugging Face’s own cloud inference endpoints. According to Phoronix, the local model server Ollama also integrates Muse Glimmer starting from version 0.32.7 on launch day; Meta announces optimized support for MLX and ExecuTorch, plus integrations with LM Studio, Unsloth, SGLang, Together AI, Fireworks AI, and OpenRouter, for the coming days.
Model inherits capabilities from the larger Muse Spark
Muse Glimmer is created through so-called logit distillation from Muse Spark, Meta’s larger cloud model, and follows the same training data in a more compact form. The smaller model inherits multi-step reasoning, reliable tool use including multimodal function calls, and error recovery for failed tasks. It processes video sequences with up to 96 frames at two frames per second without sound, handles more than 100 languages, and lets users specifically control its depth of reasoning.
On agent benchmarks such as MCP Atlas, Muse Glimmer scores 75.5 points according to Meta, against 54.2 for Gemma4-31B, while on the SWE-Bench Pro coding benchmark it scores 51.2 against 36.9 points – independently unverified. On the multimodal test Charxiv Reasoning, Muse Glimmer reaches 78.8 points; as comparison models, Meta cites Gemma4-31B and Alibaba’s Qwen3.6-27B. Meta sees use cases including local coding agents and as a judge for other AI outputs, similar to the already-launched terminal agent Muse Code, which also builds on the Muse Spark family but still depends on cloud compute.
Release lands amid Meta’s open-source strategy shift
With Muse Glimmer, Meta again positions itself as a provider of open models, after the company first switched from the open Llama series to closed Spark models in April 2026 – partly in response to criticism that Spark models trail competing products from OpenAI and Anthropic in independent tests. The new course also answers growing competitive pressure from open models out of China, such as DeepSeek, and Germany, such as Soofi S from a German research consortium.
Meta explicitly frames the open weights as a counterpoint to centralized, closed AI systems from individual providers, courting developers who want no cloud connection for cost or privacy reasons. The launch also lands during a phase of growing interest in locally running AI, as more companies decline to hand sensitive data to external server infrastructure. For freelancers and smaller firms or practices, that means an AI agent for scheduling and file organization could soon run on their own machine, with no documents leaving the premises. Meta has not yet given a rollout date for further, larger Muse Glimmer variants.
What matters now is whether independent developers confirm the benchmark figures Meta cites, and whether always-on local agent models prove trustworthy enough in practice – just weeks earlier, Meta had to disclose a security incident involving the larger predecessor model Muse Spark.
