Muse Spark Open 816B — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
816B 参数（已发布张量中为 822B，计入 1.8B 感知编码器和未绑定的嵌入），每个 token 激活 42B：256 个专家、每 token 路由 8 个，62 层分组查询注意力（16 个 KV 头，维度 64）。上下文 1,048,576 token。Meta 的模型卡列出 SGLang、vLLM、Transformers、TensorRT-LLM、Unsloth 和 llama.cpp，并称其针对两台搭载 M5 Ultra 的 Mac Studio 在 MLX 下运行做了优化。仓库为受限访问：Meta 在授权前会要求填写法定全名、出生日期和所属机构。
Muse Spark Open 816B 在 Q4_K_M 下约需 497 GB 显存——量化权重加框架开销，不含 KV 缓存。在 Apple Silicon 上，这部分来自统一内存。
按量化级别的显存与速度
计算基准：NVIDIA RTX 4090 (24 GB)。包含 8K 上下文的 KV 缓存，因此数值高于上方的主数字。
| 量化 | 显存 | 显存 | 速度（估算） | 适配 |
|---|---|---|---|---|
| Q2_K 2.63 bpw | 273.1 GB | — | 放不下 | |
| Q3_K_M 3.41 bpw | 353.3 GB | — | 放不下 | |
| Q4_K_M 4.83 bpw | 499.2 GB | — | 放不下 | |
| Q5_K_M 5.67 bpw | 585.5 GB | — | 放不下 | |
| Q6_K 6.56 bpw | 677 GB | — | 放不下 | |
| Q8_0 8.50 bpw | 876.3 GB | — | 放不下 | |
| F16 16.00 bpw | 1647 GB | — | 放不下 |
黑色标记 = NVIDIA RTX 4090 (24 GB) 上的可用显存。 估算来自内存带宽屋顶线模型，详见 方法说明页. Muse Spark Open 816B 显存计算器 →
运行 Muse Spark Open 816B
Muse Spark Open 816B is not practical on a single consumer GPU: it needs about 497.1 GB at Q4_K_M, more than any single device in our catalogue holds with headroom. How multi-GPU inference works, or rent a multi-GPU server by the hour:
如何运行 Muse Spark Open 816B
No local Ollama tag. Ollama lists muse-spark-open only as a cloud-hosted ':cloud' tag, which runs on Ollama's servers, not your machine.
规格
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- 参数量
- 816 Billion (42B active)
- 上下文窗口
- 1,048,576
- 架构
- Mixture-of-Experts (256 experts, 8 active), GQA, perception encoder
- 提供商
- Meta
- 许可证
- Muse Model License
- Availability
- Weights downloadable
- Access
- Gated: request access on Hugging Face and accept the licence first
- Practical deployment
- Multi-GPU server
- 规格量化
- Q4_K_M
- 系统内存
- 1024 GB
- 记录更新于
- 2026-10-15
Weights are downloadable and commercial use is permitted, subject to the licence’s acceptable-use terms.
Model-as-a-service operators above $20M gross revenue in any 12 months must notify Meta and sign a separate agreement; products must display "Built with Muse Spark Open".
限制与注意事项
- Gated repository: access is granted per request, and the form asks for legal name, date of birth and organisation.
- Muse Model License, not an OSI licence: products must display "Built with Muse Spark Open", and model-as-a-service operators above $20M revenue need a separate agreement with Meta.
- Not practical on any single consumer device: Meta's own minimum is 700 GB of GPU memory for the NVFP4 checkpoint and about 2 TB for BF16.
- The NVFP4 checkpoint needs Blackwell-class GPUs under vLLM.
另请参阅 Muse Glimmer 30B
质量与使用场景
评分由模型作者或独立评测方发布——衡量质量而非吞吐量，并非我们实测。
我的 GPU 能运行 Muse Spark Open 816B 吗？
Muse Spark Open 816B — 常见问题
How much VRAM does Muse Spark Open 816B need?
About 497 GB at Q4_K_M — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Muse Spark Open 816B run on an RTX 4090 (24 GB)?
No. Muse Spark Open 816B needs about 497 GB at Q4_K_M, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run Muse Spark Open 816B locally?
No local Ollama tag. Ollama lists muse-spark-open only as a cloud-hosted ':cloud' tag, which runs on Ollama's servers, not your machine. Running the published tag would send your prompts to a hosted GPU rather than your own machine.