Meta Superintelligence Labs Releases Muse Spark 1.1: A Multimodal Reasoning Model for Agentic Tasks on Meta Model API Meta超智能实验室发布Muse Spark 1.1：面向代理任务的元模型API多模态推理模型

TL;DR

• Meta Superintelligence Labs released Muse Spark 1.1, a closed, multimodal reasoning model optimized for agentic tasks with a 1-million-token context window.
• The release coincides with the Meta Model API preview, marking a shift from open weights to a hosted, metered service compatible with OpenAI and Anthropic SDKs.
• Muse Spark 1.1 demonstrates leading performance in tool use and computer use benchmarks but trails competitors like Opus 4.8 and GPT-5.5 in pure coding accuracy.
• Key architectural features include adjustable reasoning effort, active context compaction, and zero-shot delegation capabilities for parallel subagent execution.
• Pricing is set at $1.25 per million input tokens and $4.25 per million output tokens, with initial availability restricted to the US market.

Why It Matters

This release signals Meta's strategic pivot toward monetizing its frontier AI capabilities through a managed API, challenging the dominance of open-weight models in enterprise workflows that require robust agentic orchestration. By ensuring SDK compatibility with existing OpenAI and Anthropic integrations, Meta lowers the barrier for developers to integrate advanced reasoning and tool-use capabilities into their stacks without significant refactoring. The focus on long-context management and autonomous delegation positions Muse Spark 1.1 as a critical infrastructure component for complex, multi-step AI agents rather than just a conversational interface.

Technical Details

• Model Architecture & Context: Muse Spark 1.1 is a multimodal reasoning model supporting inputs of text, images, video, and documents, with a 1,000,000-token context window that the model actively compacts and manages during long sessions.
• Agentic Capabilities: The model features adjustable reasoning effort, parallel tool calling, structured output, and a delegation mechanism where it acts as a main agent planning and distributing tasks to subagents, or as a subagent executing specific jobs.
• API Compatibility: The Meta Model API is designed to be drop-in compatible with OpenAI and Anthropic formats, allowing users to switch providers by changing the base URL and API key in standard SDKs.
• Benchmark Performance: In Meta's reported evaluations, Muse Spark 1.1 led in scaled tool use (88.1 vs. 82.2 for Opus 4.8) and computer use (80.8 vs. 83.4 for GPT-5.5), but ranked third in coding benchmarks like SWE-Bench Pro (61.5 vs. 69.2 for GPT-5.5).
• Pricing Structure: The model is offered at $1.25 per million input tokens and $4.25 per million output tokens, with new accounts receiving $20 in free credits during the public preview phase.

Industry Insight

Developers should evaluate Muse Spark 1.1 specifically for agentic workflows requiring long-horizon planning and tool integration, leveraging its cost-effective pricing for high-volume token usage compared to some competitors. The closed-weight nature limits customization via fine-tuning, so organizations must assess whether the trade-off between ease of integration via standard SDKs and lack of local deployment flexibility aligns with their security and compliance requirements. As Meta enters the paid API market for frontier models, this move may accelerate the fragmentation of the AI ecosystem between open-weight self-hosted solutions and proprietary managed services, necessitating flexible architecture designs that can abstract away underlying model providers.

TL;DR

• Meta发布闭源多模态推理模型Muse Spark 1.1，主打代理任务（Agentic Tasks），支持文本、图像、视频及文档输入。
• 模型拥有100万Token上下文窗口，具备主动压缩、并行工具调用及子代理委派能力，强化长程任务执行。
• 通过Meta Model API提供，兼容OpenAI/Anthropic SDK，定价为$1.25/M输入/$4.25/M输出，目前仅限美国地区预览。
• 在工具使用和推理基准测试中表现领先，但在代码生成（SWE-Bench）和多模态理解上略逊于竞品。

为什么值得看

Meta此举标志着其从“开源权重”向“API服务化”的战略转变，为开发者提供了替代OpenAI和Anthropic的低成本选项。其独特的上下文压缩和代理委派机制，展示了处理复杂长程多步任务的新范式，对构建自动化工作流的团队具有重要参考价值。

技术解析

• 架构与能力：Muse Spark 1.1是多模态推理模型，强调“思考后回答”，推理强度可调节。支持结构化输出、并行工具调用、文件API及提示词缓存。
• 上下文管理：拥有100万Token上下文窗口，模型能主动记忆动作、检索早期信息并压缩上下文，以维持长会话效率。
• 代理机制：具备主代理（规划、委派）和子代理（执行、升级）角色。支持零样本泛化至新工具、MCP服务器及自定义技能。
• 性能基准：在MCP Atlas（88.1）和JobBench（54.7）等工具使用基准中领先；但在SWE-Bench Pro（61.5 vs 69.2）和BabyVision（76.3 vs 81.2）上排名第三。
• 集成方式：API兼容OpenAI和Anthropic格式，迁移成本低，仅需更改Base URL即可接入现有SDK。

行业启示

• API生态竞争加剧：Meta通过兼容主流SDK降低迁移门槛，迫使其他云厂商在价格和服务灵活性上做出回应。
• 代理工作流成为主流：模型从单纯的内容生成转向任务编排（Orchestration），强调工具调用、脚本编写和自主决策能力。
• 闭源与开源的分野：前沿推理模型倾向于通过API控制访问，以保障算力成本和安全性，本地部署优势可能逐渐让位于云端智能。