Meta 开源 816B 参数 Muse Spark Open 模型权重
Meta 把 816B 的编程/智能体模型权重放出来了，MoE 架构跑起来只有 42B 的速度，Mac Studio 512GB 就能装下，还给了实测数据。
Meta 在 Huggingface 上放出 Muse Spark Open 的权重，参数量 816B，主打编程和智能体任务。该开源版与 Muse Spark 1.3 共享同一底座模型，额外做了安全后训练以便在无系统护栏时独立运行，这逆转了 Meta 4 月专有闭源前沿模型的策略。模型共 62 层，每层 256 个专家网络、每个 token 路由到 8 个，因此部署需要 816B 级别的显存，但推理速度相当于 42B 模型。为解决存储需求，Meta 与 Apple 合作，M5 Ultra 提供 1.2 TB/s 统一内存，Mac Studio 最高可配 512GB。实测 3-bit 量化版在 512GB M3 Ultra 上占用约 344 GB，短提示下生成速度约 21 tok/s。
Meta released the weights of Muse Spark Open, on Huggingface, an 816B-parameter model built for coding and agentic work.
The open version shares the same base model as Muse Spark 1.3, with additional safety post-training to resist misuse as a standalone model without system guardrails, reversing Meta’s first proprietary, closed-weight frontier model strategy from April.
Each of its 62 layers holds 256 expert networks and routes every token through 8, so deployment needs the memory of an 816B-parameter model but runs at the speed of a 42B model.
That storage bill explains the Apple tie-up, because the M5 Ultra pushes 1.2 TB/s of unified memory and a Mac Studio can carry up to 512GB.
A 3-bit quant tested on a 512GB M3 Ultra took about 344 GB and ran at about 21 tok/s generation on a short prompt.
