amdenterpriseai/aim-meta-models-muse-glimmer-30b repository overview

⁠Muse Glimmer 30B

Muse Glimmer 30B is a 29.6 billion parameter dense vision-language model from Meta Superintelligence Lab, distilled from Muse Spark and built for autonomous agentic work that runs without cloud infrastructure. A 52-layer text decoder of hidden size 6656 is paired with a roughly 1.8 billion parameter ViT-G/14 perception encoder, so the model accepts interleaved text and images and can read screenshots, charts and documents alongside conversation. Three sliding-window attention layers over a 2048-token window alternate with one global-attention layer, repeated 13 times, and grouped-query attention with two key-value heads holds KV cache growth flat across the 131,072-token context. The model is trained for end-to-end task completion, reliable tool use, multi-step reasoning and failure recovery, and reasoning depth is selectable per request through a 'Reasoning strength' line in the system prompt. Output is unlike most chat models: reasoning and tool calls are written as channel-scoped messages carrying XML-style ATEM invocations rather than JSON, so both the muse_glimmer reasoning parser and the muse_glimmer tool-call parser are required and are pinned in every profile. AIM profiles target the publisher's bfloat16 checkpoint, which fits a single Instinct accelerator. No quantized profiles ship yet: the only vLLM-loadable quantizations are third-party, and no AMD Quark export of this checkpoint exists.