Meta on Thursday released Muse Spark Open, an 816-billion-parameter open-weight mixture-of-experts model with 42 billion active parameters and a one-million-token context window that the company calls the most capable U.S. open-weight model to date1,2. The omnimodal model, co-trained with Muse Code for coding agent workflows, shares the same base as Muse Spark 1.3 and includes additional safety post-training to make it more robust against misuse as a standalone model without system guardrails3,5. Its architecture uses a hybrid of 46 local attention layers and 16 global attention layers across 62 total layers, with a compressed MoE block designed to cut inter-GPU traffic and KV cache memory during long agent sessions.
Muse Spark Open is available immediately through the APIs of Baseten, Fireworks AI, Groq, Modal, Nebius Group NV, OpenRouter, and Together AI, with downloadable weights on Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face Inc., and Microsoft's Foundry Model Catalog. Baseten serves the model in NVFP4 via an OpenAI-compatible endpoint.
Benchmark positioning against Chinese rivals
Meta argues the model is on par with or beats some of the best Chinese open-weight models, depending on the benchmark. On SciCode, Meta says Muse Spark Open roughly ties Kimi K3, GLM-5.3, and Qwen3.8. Meta says the model leads all three on agentic and tool-use benchmarks including JobBench and MCP-Atlas. Meta also says Muse Spark Open beats Thinking Machines Lab's Inkling and Nvidia Corp.'s Nemotron 3 Ultra across the board in its own benchmarks, often by 20 points or more. On coding-specific evals, however, the picture is more mixed: on DeepSWE 1.1, Muse Spark Open scores 60 against Kimi K3's 68, and on Terminal-Bench 2.1, Kimi K3 scores 85.0 and GLM-5.3 scores 83.9 against Muse Spark Open's 76.4.
Meta says the model was built for coding and agentic work and excels at long-horizon tasks, complex software engineering, and multi-step reasoning. The model supports omnimodal input including text, image, video, and audio, with text output. Meta says it used Muse Spark to power Muse Code and ship updates to the Ice Cubes Mastodon client. Meta launched a closed version of Spark in April.
PyTorch toolchain and inference stack
Muse Spark Open ships with custom Helion GPU kernels integrated into the SGLang inference engine, covering operations such as RoPE, RMSNorm, and quantization. Meta says Helion megakernels fuse multiple operations into a single persistent kernel for low-latency inference, delivering 1.03–3.95x kernel-level speedups on Nvidia Corp. GB300 hardware compared to hand-tuned custom kernels. The same Helion kernels ported directly to AMD, yielding a 3% end-to-end throughput improvement on AMD MI350. The model also runs optimized on the vLLM engine, with optimization for Nvidia Corp., AMD, and Intel hardware. The full model fits on a single 4×GB300 node.
TorchTitan, a PyTorch-native training platform, adds a model definition for Muse Spark Open with native support for multi-turn supervised fine-tuning, LoRA, and reinforcement learning via TitanRL. Helion itself became part of the PyTorch Foundation in April 2026.
Local inference on Apple hardware
Meta highlighted that Muse Spark Open runs locally on Apple M5 Ultra Mac Studios, a first for desktop agentic capabilities at this parameter scale. Meta says that users with two Mac Studios equipped with Apple M5 Ultra chips and 512 GB of memory could link them and use MLX to run the model locally.
A new license with strings attached
Muse Spark Open ships under the Muse Model License rather than Meta's earlier Llama license. Meta says the license has "limited commercial terms" that allow many developers to use the model for free, but any company with more than $20 million in gross revenue during any consecutive twelve-month period must notify Meta and enter into a separate written agreement before offering the model or derivatives as a model-as-a-service business. The license also bars military and critical infrastructure deployments and requires any company using the model to prominently display "Built with Muse Spark Open" in the user interface of each product or service that incorporates it.
ANALYSIS The benchmark results position Muse Spark Open as competitive with, but not uniformly ahead of, leading Chinese open-weight models on coding tasks, while Meta claims a clearer edge on agentic and tool-use evaluations. The branding mandate and revenue-gated commercial terms mark a tighter grip on downstream use than Meta exercised with Llama, moving the Muse line closer to a source-available posture than a permissive open-weight release.