Meta open sources Muse Glimmer, a 30B local agentic model under Apache 2.0
- Meta Superintelligence Labs releases Muse Glimmer, a 30-billion-parameter model open sourced under Apache 2.0 on Hugging Face.
- The model targets always-on local agent workflows and runs on a single consumer GPU on a Mac or PC, covering local coding, function calling, and LLM-as-a-judge use cases.
- Training used a three-phase pipeline: pre-training via logit distillation from the larger Muse Spark teacher model, mid-training on longer-context agentic data, and post-training combining SFT, on-policy distillation, and reinforcement learning.
- Meta claims strong results versus Gemma4-31B and Qwen3.6-27B on agentic benchmarks like DeepSearch QA, MCP-Atlas, 𝛕-Bench, and SWE-Bench, plus tool use, multimodal input, and support for over 100 languages.
- At full precision the 30B model needs over 55GB of memory, so Meta applies quantization to fit consumer hardware while aiming to keep response latency low.
Hacker News opinions
Good to see new open weights releases from meta.
Good looking benchmarks too, which is excellent.
The least they could do, after ruthlessly bombarding my employer's servers with requests, ignoring robots.txt, scraping everything, and costing us a lot in Google Maps fees.
Meta did not abandon open source. I'd love to see a smaller distill or a MoE at this size, benchmarks look competitive as long as it's not benchmaxed, wouldn't be surprised if it is though.
Will be interesting to see how Qwen3.8 27B compares once it drops this week. Seems like dense 30B is back in fashion.
Qwen overthinking is what kills it for me, glad we have more options in this size class now.
Looking forward to trying this with llama.cpp, watching the open weights competition with high expectations.
What I really want is a model that runs on a single DGX Spark and competes with DSV4 Flash 731. Flash is already a game changer, 27b-30b just isn't smart enough for that.
DSV4 Flash 0731 already runs on an RTX 4090 24GB plus 128GB system RAM at usable tok/s with quantization.
Still needs 32-64GB of memory to run locally. A 64GB M5 MacBook Pro costs over 4k euros in Germany, a language specific model tuned for tool calling could be shrunk further.
Wow, Meta is back. Multi-token prediction makes dense models viable at speeds close to MoE but with better intelligence, and this ships pre-quantized with an MTP drafter model.
The title calling it a coding model is wrong, this is meant as a general agentic model.
The favorable comparisons to Gemma 4 and Qwen3.6 look promising.
More open weight releases grow the market for personal and small business hardware to run them, which should push cheaper hardware after years of stagnation.