Overview
<cite index="10-2,10-3">Meta Superintelligence Labs has introduced Muse Glimmer, open-sourcing the model weights under a permissive Apache 2.0 license. Muse Glimmer is a 30-billion-parameter model optimized for always-on local agent workflows.</cite> <cite index="4-2">The model was released in August 2026 and is licensed under Apache 2.0.</cite>
<cite index="4-2,4-3">Muse Glimmer is a 30-billion-parameter causal language model with a dedicated perception encoder, distilled from Muse Spark and purpose-built for autonomous agentic tasks on consumer hardware. The model integrates multi-step reasoning, reliable tool use, multimodal understanding, and failure recovery into a single model that runs locally without requiring cloud infrastructure or network access.</cite>
Architecture and Context Window
<cite index="13-10,13-11">The model has approximately 29.6 billion parameters and uses a dense Transformer design. A 1.8-billion-parameter vision encoder gives Glimmer the ability to read images, screens, charts, and documents.</cite> <cite index="12-1">The model supports a 131K context window and more than 100 languages.</cite> <cite index="4-7">Multimodal reasoning covers interpreting screenshots, charts, documents, and images alongside conversation for agentic and information-rich environments.</cite> Notably, <cite index="7-8">audio input and output are not supported.</cite>
Quantization and Hardware Fit
<cite index="17-3,17-4,17-5">Meta uses quantization techniques to compress the model's weights to approximately 4-bit precision, shrinking the language model to under 20 GB. This leaves enough headroom for the model's KV (Key-Value) cache, the perception encoder for image understanding, and the speculative decoding drafter to run simultaneously within a 24 GB or 32 GB envelope. Meta validated that this compression introduces minimum to no degradation on agentic tasks.</cite>
<cite index="19-5,19-6">Meta publishes full-precision and two 4-bit Muse Glimmer 30B checkpoints under Apache 2.0 on Hugging Face. The K-Quant-17GB release targets 24 GB hardware, while K-Quant-Dynamic targets 32 GB; Meta also ships a DFlash speculative-decoding drafter and a dedicated perception encoder.</cite>
Speculative Decoding via DFlash
A key differentiator is the bundled DFlash drafter. <cite index="17-9,17-10,17-11">The DFlash block-diffusion model predicts entire blocks of 16 tokens in a single forward pass. The main model then verifies these proposals in parallel, accepting correct tokens and correcting wrong ones. This technique lets Muse Glimmer generate text significantly faster than standard token-by-token generation while producing identical output quality.</cite>
<cite index="13-14,13-15,13-16">Meta reports 74.9 tokens per second without DFlash on an RTX 5090 and 233.4 tokens per second with DFlash, giving a 3.1× speed increase. An Apple M4 Max rises from 23.7 to 37.8 tokens per second, while an Apple M5 Max rises from 26.6 to 50.2 tokens per second.</cite>
Distribution and Toolchain Support
<cite index="2-9">The Hugging Face collection carries BF16 (Brain Float 16) weights, GGUF k-quants, ExecuTorch builds, and the DFlash drafter.</cite> <cite index="5-7">Hugging Face shipped day-zero support in Transformers, llama.cpp, vLLM, Inference Endpoints, and other libraries.</cite>
Benchmark Performance
<cite index="2-15,2-16,2-17">On published evaluations, Muse Glimmer leads on MCP (Model Context Protocol) Atlas at 75.5, DeepSearch QA at 74.6, Gaia2 at 43.3, and SWE-Bench Pro at 51.2. Reasoning scores include AIME 2026 at 94.7, IFBench at 77.0, and AA-LCR at 80.0.</cite> <cite index="2-18,2-19">Qwen3.6-27B leads on OSWorld-Verified (75.6 versus 65.9), TerminalBench 2.1 at 60.7, and SWE-Bench Verified at 77.2.</cite>
Safety and Licensing Context
<cite index="2-24,2-25">Meta states the model does not meet the Frontier AI definition in its Advanced AI Scaling Framework, and rates chemical/biological, cyber, and loss-of-control risk at moderate or lower.</cite> <cite index="9-1">Meta co-founder and CEO Mark Zuckerberg wrote that the weights for Muse Glimmer, "a great 30B parameter dense model that can run locally," are now open.</cite> <cite index="8-3,8-4">The release marks Meta's second open-weighted Muse model in five months; where April's Muse Spark stayed closed and July's Muse Spark 1.1 shipped only as a hosted API, Muse Glimmer landed on Hugging Face with weights under Apache 2.0, small enough to run on a single consumer GPU.</cite>