Muse Spark Open

The OpenLM model card describes Muse Spark Open as an 816B total, 42B active open-weights model with a 1M-token context and native multimodality, adding extra safety post-training so it can stand alone without system guardrails. It targets multi-step reasoning, end-to-end agentic task completion, coding agents, tool use, multimodal reasoning, synthetic data generation, and LLM-as-a-judge evaluation. Checkpoints ship in NVFP4 and BF16. On an OpenHands multi-turn agentic coding replay using NVFP4 on eight NVIDIA B300 GPUs, SGLang reaches 14.8K total tokens per second per GPU at concurrency 1 and 80.3K at concurrency 128, with sequence parallelism, multimodal preprocessing optimizations, and DFlash V2 speculative decoding.

Meta AI Research released open weights for Muse Spark Open, which it calls the most capable US open-weight model to date. The model is natively multimodal with 816B total parameters, 42B active, and a 1M-token context window, built for coding and agentic work including long-horizon reasoning and complex software engineering. Weights are downloadable from Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face, and Microsoft Foundry, and the model runs on SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp, and Unsloth with silicon optimizations. Meta also optimized it for Apple M5 Ultra Mac Studios, where two machines with 512 GB each hold the model via unified memory.

This analysis reports that Muse Spark Open leads US open-weight models on the Artificial Analysis Intelligence Index and on tool-calling agents, factual reliability, and image questions, while trailing Kimi K3 and GLM 5.3 on coding agents. The architecture is a 256-expert, top-8 MoE with 62 layers, split between sliding-window and full attention layers. It adds that hybrid attention cuts the KV cache by 3.7 times at 128K context, that BF16 weights total 1.6 TB while NVFP4 quantization is 530 GB and fits a single 8-GPU or 4x GB300 node, and that hosted APIs deliver about 102 tokens per second with 0.67-second time to first token.

Unsloth's guide covers running Meta's Muse Spark Open locally as an 816B-parameter MoE model with 42.7B active, supporting text, image, video, and audio inputs with text output and a 1M-token context. Dynamic 2-bit GGUF quants fit in 256 GB of memory, while 8-bit requires roughly 900 GB, against about 1.63 TB for full-precision BF16 weights. The recommended starting point is the Dynamic 2-bit Muse-Spark-Open-816B-A42B GGUFs, and llama.cpp can offload to disk when memory is short. Because the model is a reasoning model, it supports low through xhigh reasoning efforts with temperature 1.0, top p 1.0, and top k disabled at a 1,048,576 maximum context.

The Hugging Face repository for Muse-Spark-Open-816B-A42B publishes the model under the Muse Model License with a release date of October 15, 2026 and copyright to Meta Platforms. The license grants free use, modification, fine-tuning, distillation, merging, redistribution, sublicensing, and commercial sale of the weights. Users must display "Built with Muse Spark Open" in the interface or documentation of any product using the software or its derivatives. It also imposes an attribution-style condition on Model as a Service businesses, requiring a copyright notice, the license, and usage policy to accompany copies and derivatives.