muse glimmer has better taste than expected on voxel worlds. both one-shot, generated using my GX10 structure isn't perfect and it's less convincing than ds4 flash or qwen 3.6 but it's good enough!

Uncensored Meta Muse Glimmer 30B model run locally 10.7 GB on 5090. - Claim on the card: 0/300 measured refusals. - 131k context. - image understanding Agent out. - compact vision projector + DFlash drafter included - claimed 0 true refusals on a 450-prompt harmful suite -

Today we published a technical report on Muse Glimmer, our 30B open-weight model built for agentic tasks on consumer hardware. It runs at practical speeds on a Mac or PC with a single consumer GPU at 4-bit precision, and leads comparably-sized open models on tool use and multi-step planning. Read about the model architecture, distillation recipe, quantization and more:

In the past 6 months... - Muse Spark - Muse Image - Muse Video - Muse Spark 1.1 - Meta Model API - Muse Glimmer - Muse Spark 1.2 - Muse Code - Muse Spark 1.3 - Muse ...and we’re just getting started. 🚀

Fine-tune Meta AI's Muse Glimmer 100% locally, using Lightning's new studio template and Unsloth. QLoRA, 4-bit quantization, and embedding offload let you run it on a 24 GB GPU: fine-tune the vision tower, language model, or both, then save as LoRA adapters or a merged model ready for deployment. Clone it and start building →

You can now fine-tune Meta Muse Glimmer 30B for free! 🔥 Our free notebook also supports GRPO RL training. Unsloth trains Muse Glimmer 1.5× faster with 50% less VRAM vs FA2 setups. Train locally with 24GB VRAM. Guide: Notebooks:

I shipped a fine-tuning tutorial for Muse Glimmer 30B on AI2's MolmoWeb dataset with TRL > but merve, this model is sota on ScreenSpot-Pro? yes, but it doesn't work well with ambiguous prompts of MolmoWeb that are close to how you interact with computer, e.g. "jump to nutrition facts" I compared model fine-tuned on MolmoWeb format against base model zero-shot outputs converted to MolmoWeb format (coordinates on 100) I found that base model has 13% click accuracy + 35.0% within 5% diagonal while fine-tuned model has 41% + 68% within 5% diagonal

free Kaggle dual T4s GPUs Running Meta’s Muse Glimmer 30B Q6_K_XL. - 130,000 token context - 26.5 GB VRAM - Prefill: 265 t/s - Decode: 9 t/s - Cost: $0 Aggressive 16:1 GQA. Unquantized KV cache for 130k tokens is only 2 GB.