Meta Releases Open Model Muse Glimmer With 29.6B Parameters
Muse Glimmer shows a local, open model at 29.6B parameters outperforming comparable closed and open rivals on benchmarks, narrowing the gap between device-runnable AI and cloud models.
Reporting from 1 source: GIGAZINE.
Meta released Muse Glimmer, an open AI model with 29.6 billion parameters, on August 10, 2026. The dense model handles text and image inputs. Benchmarks show it outperforming Google's Gemma 4 31B across many tests. It supports the DFlash speculative decoding method for faster local processing.
Meta positioned Muse Glimmer as a device-runnable model with competitive performance. The company published benchmark tables comparing it against Gemma 4 31B and Qwen3.6-27B, with Muse Glimmer leading in most tests.
Third-party evaluator Artificial Analysis also tested the model. Its composite index places Muse Glimmer above the closed Claude Haiku 4.5 and close to Gemini 3.5 Flash-Lite. Meta described the result as a large jump over the Llama 4 series from 2025.
The model uses DFlash, a speculative decoding method that generates drafts with a lightweight diffusion model. On an NVIDIA GeForce RTX 5090, DFlash raises decode speed by 3.1 times to 233 tokens per second. Weights fit in 24GB of VRAM, and the license is Apache License 2.0. LM Studio and Ollama already support it.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.