Muse Spark Open: Meta’s 816B Open-Weight Multimodal Agent Model

Meta released Muse Spark Open on 15 October 2026. It is the open-weight release of the Muse Spark model family, built for coding agents and long-horizon work. The headline is the combination of sparse compute, native multimodal input, and a very long context window—not a claim that all 816B parameters run for every token.

The language backbone is a 62-layer mixture-of-experts Transformer. Each token is routed to 8 of 256 experts in each MoE layer, so only a fraction of expert parameters is active for a token. Sparse activation reduces per-token computation, but it does not eliminate the need to store or distribute the full checkpoint during inference.

Muse Spark Open accepts text, images, audio and video and produces text. A shared perception encoder handles images and video frames, while a separate compact audio encoder turns waveform segments into audio tokens. The model can therefore reason over mixed evidence such as code, screenshots, video frames and audio, which is useful for debugging workflows that cannot be described as text alone.

The model is designed for multi-step workflows: inspect a codebase, plan a change, invoke tools, observe failures, and revise the result. This differs from optimizing only for a single code completion. Meta describes internal demonstrations that used Muse Code to modify and build an open-source iOS app and to edit video in DaVinci Resolve. These are vendor demonstrations, not independent benchmark results.

Inference partners report optimizations for the model’s sparse routing and long context. vLLM documents DSpark speculative decoding, which proposes several draft tokens using block diffusion and verifies them with the main model. These speed claims depend on hardware, batch size, context length, acceptance rate and software version; do not treat them as universal latency guarantees.

Meta’s launch page points readers to an evaluation methodology report and detailed model card rather than supplying a complete, independently reproducible benchmark table in the announcement itself. This article therefore focuses on disclosed architecture, deployment and implementation details. Performance statements from Meta, NVIDIA and vLLM are attributed to those sources and should not be interpreted as neutral cross-model rankings.

The open-weight release includes additional safety post-training, and Meta says it assessed the model under its own scaling framework. That is not a substitute for application-specific safeguards. The Muse Model License allows broad use subject to its stated conditions and is not interchangeable with Apache 2.0. Teams should review the actual license and model card before deployment.

Muse Spark Open is notable because it combines open weights, sparse 816B-scale MoE architecture, native multimodal input and a 1M context window with day-one support across multiple inference stacks. Its practical cost is determined not just by the active parameter count but by checkpoint precision, resident memory, inter-GPU communication, KV cache, speculative decoding and the agent harness around it.