Muse Spark Open
An open-source Mixture-of-Experts large language model with 816B parameters, supporting NVFP4 4-bit quantization, suited to distributed inference deployment.
Core Facts
Timeline (last 90 days)
On the MCP-Atlas (tool calling) benchmark, Muse Spark Open scores 89.2 and GLM 5.3 scores 84.2.
Meta recommends using two M5 Ultra Mac Studios each with 512 GB of unified memory to run Muse Spark Open.
On the Intelligence Index, Muse Spark Open scores 39 and GLM 5.3 scores 45.
On DeepSWE 1.1, Muse Spark Open scores 60 and Kimi K3 scores 68.
On Terminal-Bench 2.1, Muse Spark Open scores 76.4, while Kimi K3 scores 85.0.
Meta released the Muse Spark Open open-source large model, which uses a MoE architecture, has a total parameter count of 816B and 42B activated parameters, supports a 1M token context, and is released in the form of open weights.
Muse Spark Open's results on MCP-Atlas mean that it is not just a general-purpose chat model, but has been specifically optimized for Agent workflows.
Muse Spark Open leads existing open-source models in the MCP-Atlas evaluation.
Meta released an open-weight model called Muse Spark Open, which can run locally on a Mac.
Meta released the open weights of Muse Spark Open on October 15, calling it officially the strongest US open-weight model to date.
40 more timeline events
All Facts (20)
Muse Spark can be used as a backend via OpenCode Go.
85%Partially VerifiedMuse Wonder supports a 1 million Token context window.
75%UnverifiedThe actual retrieval accuracy of Muse Spark Open under a 1M context window has not yet been verified and requires developers to test it themselves.
70%UnverifiedMuse Spark Open uses an MoE architecture, with 816B total parameters and 42B activated parameters.
70%UnverifiedHugging Face calculates that the KV cache under BF16 is about 250 KB per token, and a full 1-million-token context requires about 266 GB.
60%UnverifiedMuse Spark Open has 816 billion total parameters, activates 42 billion parameters per token, and has a context window of 1 million tokens.
60%UnverifiedMuse Spark Open is the first model of its kind that can run locally on two Mac Studios with Apple M5 Ultra chips with the help of the MLX framework.
60%UnverifiedMuse Spark Open uses the Muse Model License.
60%UnverifiedMuse Spark Open is positioned as the most powerful open-weight model in the United States to date.
60%UnverifiedMuse Spark Open supports a 1 million token context window.
60%UnverifiedMuse Spark Open's weight download channels include Amazon SageMaker AI, Google Cloud Model Garden, Hugging Face and Microsoft Foundry Models.
60%UnverifiedMuse Spark Open can be called on the day of release through platforms such as Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI.
60%UnverifiedMuse Spark Open, the open-weight inference model released by Meta, has officially been integrated into AI Gateway.
60%UnverifiedOn the Intelligence Index, Muse Spark Open scores 39 and GLM 5.3 scores 45.
50%UnverifiedOn Terminal-Bench 2.1, Muse Spark Open scores 76.4, while Kimi K3 scores 85.0.
50%UnverifiedOn DeepSWE 1.1, Muse Spark Open scores 60 and Kimi K3 scores 68.
50%UnverifiedOn the MCP-Atlas (tool calling) benchmark, Muse Spark Open scores 89.2 and GLM 5.3 scores 84.2.
50%UnverifiedMeta recommends using two M5 Ultra Mac Studios each with 512 GB of unified memory to run Muse Spark Open.
50%UnverifiedMuse Spark Open's results on MCP-Atlas mean that it is not just a general-purpose chat model, but has been specifically optimized for Agent workflows.
50%UnverifiedMuse Spark Open leads existing open-source models in the MCP-Atlas evaluation.
50%