Muse Glimmer 30B vs Muse Spark 1.2

Muse Glimmer 30B is a dense vision language model from Meta built for long-horizon agentic work on local hardware. The architecture pairs a 52-layer causal text decoder with a roughly 1.8B parameter ViT-G/14 perception encoder for about 29.6 billion parameters in total, and it accepts interleaved text and image input so an agent can interpret screenshots, charts, and documents alongside conversation. The decoder uses grouped-query attention with 32 query heads and 2 key-value heads, a repeating pattern of three sliding-window local attention layers followed by one global layer, SwiGLU feed-forward blocks, and rotary position embeddings applied on the local layers, supporting a trained context of 131,072 tokens.

Meta describes the model as distilled from the larger Muse Spark and trained and evaluated around agentic behavior: end-to-end task completion, schema-accurate tool calling, multi-step reasoning across extended workflows, and recovery when a tool call returns an unexpected result. Reasoning effort is selectable across low, medium, high, and xhigh settings, and the model emits channel-scoped reasoning traces together with XML style tool calls rather than JSON, which requires parsers specific to this family. A companion block-diffusion drafter head predicts blocks of 16 tokens per forward pass for speculative decoding, with the main model verifying the proposals in parallel.

Muse Spark 1.2 is a proprietary multimodal reasoning model from Meta Superintelligence Labs, released as a coding-focused update to Muse Spark 1.1. It accepts text, images, video, audio, and PDF documents and returns text, with a context window of roughly one million tokens that allows whole repositories, long documents, and extended agent trajectories to be held in a single request. The model thinks before answering, and the amount of reasoning effort it spends is configurable per request. Alongside its visual and document understanding, it supports structured output and parallel function calling, and it is designed to operate either as a planning agent that delegates work or as a subagent executing tasks in parallel.

Training for version 1.2 scaled up compute on coding tasks and widened the diversity of training environments, concentrating on long-horizon work such as whole-repository generation, large end-to-end projects, and automated research. Part of the training data was self-generated, with Muse Spark 1.1 producing coding environments and instruction-following templates and grading candidate solutions against them. The model was co-trained with the Muse Code terminal agent, incorporating rejection-sampled harness trajectories and that toolset. Meta reports 82.9 percent on Terminal-Bench 2.1, an improvement of 6.7 points over Muse Spark 1.1. Multimodal use cases documented for the family include visual-to-code generation and detailed image and video captioning.