Muse Spark 1.3 是 Meta 推出的对话大模型,发布于 2026-09。Muse Spark 1.3 is a Chat Model released by Meta, launched on 2026-09.
Meta Superintelligence Labs 2026-09-02 发布的闭源多模态推理模型,主攻长周期智能体与编码:1M 上下文、131K 输出,输入 $1.25 / 输出 $4.25 每百万 tokens(缓存读 $0.15)。Meta 自测 DeepSWE v1.1 75.4、Terminal-Bench 2.1 88.8、MRCR 512K-1M 长文档召回 98.1;官方开放 xhigh 推理档,max 档待安全测试后放行。另设「贡献者」低价端点 $0.10/$0.20,代价是 Meta 可用你的流量训练模型。适合整仓检索、长文档一次性读完的场景。Meta Superintelligence Labs' closed multimodal reasoning model, released 2026-09-02 for long-horizon agents and coding: 1M context, 131K output, at $1.25 input / $4.25 output per million tokens (cache reads $0.15). Meta reports 75.4 on DeepSWE v1.1, 88.8 on Terminal-Bench 2.1 and 98.1 long-document recall on MRCR 512K-1M. The xhigh reasoning tier ships today; max remains gated pending safety testing. A 'contributor' endpoint at $0.10/$0.20 trades data rights for a steep discount. Best suited to whole-repo retrieval and single-pass long-document reading.
1M 长文档召回 98.1%、多模态含音频、视频与 PDF、编码智能体表现第一梯队98.1% recall across 1M-token documents、Audio, video and PDF input、First-tier agentic coding
注意事项
输出冗长,实际成本高于标价、max 推理档尚未开放、闭源,开放权重无时间表Verbose output, real cost above sticker price、Max reasoning tier not yet available、Closed weights with no open-release date
Muse Spark 1.3 目前以付费 API / 平台调用为主,可关注官方是否提供免费试用额度。Muse Spark 1.3 is mainly available via paid API/platform calls; watch for official free trial quotas.
Muse Spark 1.3 的上下文窗口有多大?How large is Muse Spark 1.3's context window?
Muse Spark 1.3 的上下文窗口为 1M,适合处理长文档与多轮复杂对话。Muse Spark 1.3's context window is 1M, suitable for long documents and multi-turn complex dialogue.
Muse Spark 1.3 是开源还是闭源?Is Muse Spark 1.3 open-source or closed-source?
闭源模型,主要通过官方 API 或平台使用。Closed-source model, mainly available via official API or platform.
Muse Spark 1.3 适合用来做什么?What is Muse Spark 1.3 good for?
Muse Spark 1.3 的主要能力包括 chat、vision、video、audio、pdf、function、reasoning、agent,可用于相关场景的创作、分析与自动化任务。Muse Spark 1.3's main capabilities include chat、vision、video、audio、pdf、function、reasoning、agent, usable for creation, analysis and automation in related scenarios.