Muse published a technical report for Muse Glimmer, a 30B open-weight model built for agentic tasks on consumer hardware. The model runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision. Muse claims it leads comparably-sized open models on tool use.
Meta released Muse Glimmer, a 30B-parameter open-weight model under Apache 2.0, optimized for local agentic workflows. It runs on consumer GPUs with 24GB memory and under 20GB download, supporting planning, tool use, and self-correction. The release aligns with Zuckerberg's manifesto on decentralized AI control.
Meta released Muse Glimmer, a 30B-parameter open-weight model under Apache 2.0, optimized for local agentic workflows. It runs on consumer GPUs with 24GB memory and under 20GB download, supporting planning, tool use, and self-correction. The release aligns with Zuckerberg's manifesto on decentralized AI control.
A Reddit user describes running Muse Glimmer (30B) locally as a llamafile with the Hermes coding agent, which autonomously fixed a real bug in Mozilla AI's Otari gateway and opened draft PR #727. No API keys or cloud were used, suggesting ~30B local models are becoming viable for real agentic dev workflows.
A Reddit user describes running Muse Glimmer (30B) locally as a llamafile with the Hermes coding agent, which autonomously fixed a real bug in Mozilla AI's Otari gateway and opened draft PR #727. No API keys or cloud were used, suggesting ~30B local models are becoming viable for real agentic dev workflows.
Users report that the new Muse Glimmer model refuses to generate even benign code, such as mouse-control scripts, citing potential misuse. The refusals occur in local setups like Unsloth Q8 on dual 3090s within Kilo Code, raising concerns about over-censorship in a supposedly open model.
Users report that the new Muse Glimmer model refuses to generate even benign code, such as mouse-control scripts, citing potential misuse. The refusals occur in local setups like Unsloth Q8 on dual 3090s within Kilo Code, raising concerns about over-censorship in a supposedly open model.
A Reddit user ran eleven matched on/off MTP tests across Gemma 4 and Qwen3.6, finding consistent speedups of 1.65x to 2.54x with no measurable accuracy loss. Meta's Muse Glimmer drafter underperformed, slowing a 7900 XTX by 9% and keeping only 24.55% of drafted tokens, contradicting Meta's claimed 3.1x speedup on an RTX 5090.
A Reddit user ran eleven matched on/off MTP tests across Gemma 4 and Qwen3.6, finding consistent speedups of 1.65x to 2.54x with no measurable accuracy loss. Meta's Muse Glimmer drafter underperformed, slowing a 7900 XTX by 9% and keeping only 24.55% of drafted tokens, contradicting Meta's claimed 3.1x speedup on an RTX 5090.
Meta announced a renewed focus on open-weight LLMs, releasing the open model Muse Glimmer and promising to open the weights for its more powerful Muse Spark 1.2 within weeks. CEO Mark Zuckerberg published a 6,500-word manifesto outlining his AI philosophy, drawing criticism for its tone.
Meta announced a renewed focus on open-weight LLMs, releasing the open model Muse Glimmer and promising to open the weights for its more powerful Muse Spark 1.2 within weeks. CEO Mark Zuckerberg published a 6,500-word manifesto outlining his AI philosophy, drawing criticism for its tone.