What Is Muse Spark Open?Muse Spark Open 是什么？
Muse Spark Open is Meta's open-weight release of its Muse Spark model family, published October 15, 2026 — three days after the closed flagship Muse Wonder (October 12), and the lab's return to publishing frontier-scale weights. The headline numbers: 816 billion total parameters with 42 billion active per token (256 experts, top-8 routing), a 1M-token context window, and native multimodal input — text, images, video and audio in, text out. Meta describes it as trained specifically for agentic coding: reading a codebase, writing and debugging code, and calling tools. It is the full-size sibling of Muse Glimmer, the small dense model Meta released in August, and shares its tokenizer.Muse Spark Open 是 Meta 2026 年 10 月 15 日发布的 Muse Spark 模型家族开权重版——在闭源旗舰 Muse Wonder（10 月 12 日）发布三天后到来，是 Meta 时隔一年多再次放出前沿级权重。核心数字：816B 总参数、每 token 激活 42B（256 个专家、top-8 路由）、100 万上下文，原生多模态输入——文本/图像/视频/音频进、文本出。Meta 称其专为 Agent 编程训练：读代码库、写代码、调工具。它是 8 月发布的小型稠密模型 Muse Glimmer 的全尺寸兄弟版，共用同一 tokenizer。
The weights are free to download from Hugging Face under the Muse Model License — but running them is another story: the model card lists ~700 GB of VRAM minimum for the NVFP4 build (533 GB across 127 shards; BF16 is 1.64 TB across 65 shards), targeting Blackwell GPUs. The realistic path for most teams is a hosted endpoint: $1.25 input / $4.25 output per million tokens on OpenRouter and Together AI (cached input $0.15), with seven launch providers: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Prices per the official site.权重在 Muse Model License 下可从 Hugging Face 免费下载——但跑起来是另一回事：模型卡要求 NVFP4 版最低约 700GB 显存（NVFP4 533GB/127 分片；BF16 1.64TB/65 分片），面向 Blackwell GPU。大多数团队的现实路径是第三方托管：$1.25 输入/$4.25 输出每百万 token（OpenRouter、Together AI；缓存输入 $0.15），首批 7 家上线：Baseten、Fireworks AI、Groq、Modal、Nebius、OpenRouter、Together AI。价格以官网为准。
Muse Spark Open is the most consequential open-weight release of 2026: an 816B American frontier model you can actually download, at a moment when the open frontier had been ceded to Chinese labs for over a year. The catch is honest arithmetic — 700 GB of VRAM means almost nobody self-hosts, and the $1.25/$4.25 hosted pricing makes the endpoint the only sane choice under billions of tokens a month. The license is genuinely permissive for most teams ($20M revenue threshold only bites if you sell hosted API access; embedding it in your own product is exempt), and the safety card reads well — StrongREJECT v2 adaptive-attack score of 0.5 where rivals score 62.9–86.4. The honest tradeoff: on pure coding benchmarks it trails Kimi K3, GLM 5.3 and Qwen3.8 in Meta's own numbers, so pick this for agentic tool-use workloads and data-sovereignty use cases, not raw code completion. Verdict: if you need open weights with a US license pedigree and a real hosting market on day one, this is the default — start on OpenRouter with the hosted endpoint and only build the GPU cluster if the bill justifies it.
Muse Spark Open 是 2026 年最有分量的开权重发布：816B 美系前沿模型终于能下载了——而在此之前，开源前沿已经让给中国厂商一年多。代价是诚实的算术：700GB 显存意味着几乎没人能自托管，月用量不到几十亿 token 的话，$1.25/$4.25 的托管价是唯一理性的选择。许可证对大多数团队真正友好（$20M 营收门槛只针对卖托管 API 访问的场景，把模型嵌进自己产品里豁免），安全卡也不错——StrongREJECT v2 自适应攻击得分 0.5，对手们在 62.9–86.4。真实代价：纯编程基准上它在 Meta 自己的数字里落后 Kimi K3、GLM 5.3、Qwen3.8，所以选它是为 Agent 工具调用和数据主权场景，不是为原始代码补全。结论：如果要开权重+美国许可证背书+首日就有真实托管市场，它是默认选项——先在 OpenRouter 上用托管端点跑起来，账单撑得住再考虑搭 GPU 集群。
— 中国吧 AI Tools Hub Editorial Team
Who Should Use Muse Spark Open?谁适合使用 Muse Spark Open？
✓ Good Fit For适合以下场景
- Teams that need open weights with a permissive commercial license需要开权重+宽松商用许可的团队
- Agent/coding workloads with heavy tool use and long contexts重工具调用、长上下文的 Agent/编程场景
- Data-sovereignty cases where the model must run in your own infra数据主权场景：模型必须跑在自己的基础设施里
✕ Not Ideal For不适合以下场景
- Anyone expecting to self-host on a laptop — ~700 GB VRAM minimum想在笔记本上自托管的人——最低约 700GB 显存
- Pure code-completion speedruns — rivals lead on Meta's own coding benchmarks纯代码补全竞速——Meta 自己的编程基准上对手领先
- Sellers of hosted API access above $20M revenue — needs a separate Meta agreement年营收超 $20M 且卖托管 API 访问的——需与 Meta 另签协议
Getting Started with Muse Spark OpenMuse Spark Open 快速开始
1. Grab the weights: Hugging Face (free download, access form) — NVFP4 build for Blackwell GPUs, or BF16 for max fidelity拿权重：Hugging Face（免费下载，需填申请表）——Blackwell GPU 用 NVFP4 版，要最高保真用 BF16 版2. Skip the hardware: create an API key on OpenRouter or Together AI — $1.25 input / $4.25 output per MTok (prices per official site)跳过硬件：去 OpenRouter 或 Together AI 创建 API key——$1.25 输入/$4.25 输出每百万 token（价格以官网为准）3. Point your OpenAI-compatible client at it and set a long context — 1M tokens fits whole repos for agentic codingOpenAI 兼容客户端直接指过去，把上下文拉满——100 万 token 能把整个代码库塞进去做 Agent 编程Key Features核心功能
-
816B MoE, 42B Active816B MoE、42B 激活 — 816 billion total parameters with 42 billion active per token (256 experts, top-8 routing) — fine-grained DeepSeek-style routing that keeps inference cheaper than a dense model of the same size.816B 总参数、每 token 激活 42B（256 个专家、top-8 路由）——DeepSeek 风格细粒度路由，推理成本远低于同体量稠密模型。
-
1M-Token Context100 万上下文 — A million-token context window for whole codebases and long multi-turn agent loops — trained on data from 100+ languages.100 万 token 上下文——整个代码库、超长多轮 Agent 链路一次吃下；训练数据覆盖 100+ 语言。
-
Free Downloadable Weights可免费下载的权重 — BF16 (1.64 TB, 65 shards) and NVFP4 (533 GB, 127 shards) checkpoints on Hugging Face, with mirrors on AWS, Google Cloud and Microsoft; ~700 GB VRAM minimum to self-host.Hugging Face 上 BF16（1.64TB/65 分片）与 NVFP4（533GB/127 分片）双版本，AWS/Google Cloud/微软镜像；自托管最低约 700GB 显存。
-
Hosted API from $1.25/$4.25托管 API $1.25/$4.25 起 — OpenRouter and Together AI list $1.25 input / $4.25 output per MTok (cached input $0.15); seven providers at launch: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, Together AI. Prices per official site.OpenRouter、Together AI 挂牌 $1.25 输入/$4.25 输出每百万 token（缓存输入 $0.15）；首批 7 家：Baseten、Fireworks AI、Groq、Modal、Nebius、OpenRouter、Together AI。价格以官网为准。
-
Muse Model LicenseMuse Model License — Free use, fine-tuning, distillation and redistribution for most developers; only hosted Model-as-a-Service above $20M revenue needs a separate Meta agreement; products must show a 'Built with Muse Spark Open' label.大多数开发者可免费使用、微调、蒸馏、再分发；只有年营收超 $20M 的托管 MaaS 需与 Meta 另签协议；产品须标注"Built with Muse Spark Open"。
Pros & Cons优缺点
✓ Good Fit For适合以下场景
- Frontier-scale open weights with a genuinely permissive commercial license前沿级开权重，商用许可真正宽松
- Real hosting market on day one — 7 providers, $1.25/$4.25 pricing首日就有真实托管市场——7 家供应商、$1.25/$4.25 定价
- Strong safety card: StrongREJECT v2 0.5 vs rivals' 62.9–86.4 (vendor-reported)安全卡漂亮：StrongREJECT v2 0.5，对手 62.9–86.4（厂商自报）
- Agent tool-use leadership in Meta's own benchmarksMeta 自己的基准里 Agent 工具调用领先
✕ Not Ideal For不适合以下场景
- Self-hosting is a fantasy for most — ~700 GB VRAM minimum自托管对大多数人是幻想——最低约 700GB 显存
- Trails Kimi K3 / GLM 5.3 / Qwen3.8 on coding tests (Meta's own numbers)编程测试落后 Kimi K3/GLM 5.3/Qwen3.8（Meta 自己的数字）
- Mandatory 'Built with Muse Spark Open' product label产品必须标注"Built with Muse Spark Open"
Use Cases应用场景
Popular ways people use Muse Spark Open.人们使用 Muse Spark Open 的常见场景：
💻 Agentic Coding on Whole Repos整库级 Agent 编程
1M context swallows entire codebases — the model's stated training target is reading, writing and debugging code with tool calls.100 万上下文把整个代码库吞进去——模型的训练目标就是读、写、调代码加工具调用。
🔧 Private AI Infrastructure私有 AI 基础设施
Data-sovereignty teams run the weights in their own cloud (or two M5 Ultra Mac Studios via MLX, ~25 tok/s) instead of sending data to a vendor API.数据主权团队在自己的云里跑权重（或两台 M5 Ultra Mac Studio 走 MLX，约 25 tokens/s），而不是把数据送到供应商 API。
🧪 Fine-Tuning & Distillation微调与蒸馏
The license explicitly permits fine-tuning, distillation and redistribution — a legal foundation for derivative models.许可证明确允许微调、蒸馏、再分发——做衍生模型的法律地基。
💸 Cheap Frontier Inference便宜的前沿推理
Hosted at $1.25/$4.25 per MTok with $0.15 cached input (prices per official site) — competitive against hosted-only flagships for high-volume agent loops.托管价 $1.25/$4.25 每百万 token、缓存输入 $0.15（价格以官网为准）——大用量 Agent 链路上对纯托管旗舰有竞争力。
Similar AI Tools相似 AI 工具
If Muse Spark Open doesn't fit your needs, here are other popular AI tools you might consider — for a comparable open-weight flagship see our full review: MiniMax M3.1 Flash. The closed frontier sibling is covered separately — see Nano Banana 2.1 for this batch's image-generation flagship.如果 Muse Spark Open 不符合你的需求，可以看看这些热门 AI 工具——同类开权重旗舰可参考本站完整评测：MiniMax M3.1 Flash。本批次的图像生成旗舰是 Nano Banana 2.1。