最新模型
Meta Releases Open Weights for Muse Spark, an 816B Multimodal Model in BF16 and NVFP4
Meta released Muse Spark Open on October 15. Its MoE architecture activates about 42B parameters per token and supports a million-token context. Hugging Face has published deployment instructions and Mac quantization benchmarks, though the full weights and long context still require substantial memory.
On October 15, Meta released the open weights for Muse Spark Open, integrating text, image, audio, and video inputs into a single model with text output. It is positioned for software development and long-running agent tasks, including cross-file edits, tool calls, and error recovery. Meta also lists inference options including SGLang, vLLM, Transformers, MLX, and llama.cpp, allowing teams to deploy and tune the model themselves. Meta announcement
Muse Spark Open has about 816B total parameters, with about 42B activated per token. The language model has 62 layers, and each MoE block has 256 experts, of which the router selects 8 at a time. Its attention uses 128 query heads and 16 KV heads, with QK normalization. This design reduces the number of parameters involved in computation per token, but the remaining expert weights still need to be stored, so deployment requirements cannot be estimated as if this were a 42B model. Official model card
The release includes BF16 and NVFP4 versions. Hugging Face reports that the BF16 weight files are about 1.5 TB. It also tested an oQ3e-quantized version on an M3 Ultra with 512 GB of memory; the model used about 344 GB. With an input of 1,024 tokens and 128 generated tokens, prompt processing measured about 225 tokens per second, and generation about 21 tokens per second. These results offer a concrete reference for local inference, but do not represent performance with a million-token context, multiple concurrent users, or a complete agent workflow. Hugging Face deployment and benchmarks
For engineering teams, this release adds another self-hostable multimodal agent model option and could allow screenshot analysis, code edits, and tool execution to share a single model. During evaluation, teams should calculate requirements for weights, KV cache, and runtime buffers separately, and measure tool-call success rates and long-task completion rates after quantization. The official model card lists xhigh as the default inference effort; latency and output token costs should also be included in comparisons. Chinese code comments, document comprehension, and Traditional Chinese instruction following still require separate testing. Model settings and limitations
The model is licensed under the Muse Model License, which includes usage restrictions and conditions for some commercial uses. Teams should check how these apply to their service before adopting it. Areas to watch include stable support across inference frameworks and reproducible benchmarks across different hardware, quantization precisions, and context lengths. Meta licensing information