Meta 发布 Muse Spark 1.1：面向智能体任务的多模态推理模型

Today, Meta Superintelligence Labs released Muse Spark 1.1. Alongside it, Meta opened a public preview of the Meta Model API. That second part is the structural change. Meta’s models previously reached developers mainly as open weights. Muse Spark 1.1 is closed, hosted, and metered per token. So the question is narrow. Where does it belong in a stack you already run?

What is Muse Spark 1.1?

Meta describes it as a multimodal reasoning model built for agentic tasks. Reported gains over the first Muse Spark sit in tool use, computer use, coding, and multimodal understanding. The context window is 1,000,000 tokens. Meta Model API docs list 1,048,576.

Because it is a reasoning model, so it thinks before answering. Furthermore, this reasoning effort is adjustable per request. Inputs span text, images, video, and documents; output is text. The API also exposes structured output, parallel tool calling, a Files API, and prompt caching. Adding a web_search tool to a Responses API call returns cited answers.

Access splits two ways. Consumers get it free in ‘Thinking’ mode in the Meta AI app and on meta.ai. Developers pay $1.25 per million input tokens and $4.25 per million output tokens. New accounts get $20 in free credits. Initial launch post describes the public preview as US-only, with no EU access yet.

Performance

With the spec in place, the numbers explain the positioning. To illustrate this, Meta published a launch table, and the table splits cleanly.

BenchmarkTestsMuse Spark 1.1Opus 4.8 (max)GPT-5.5 (xhigh)Gemini 3.1 Pro (high)MCP AtlasScaled tool use88.182.275.378.2JobBenchProfessional tool use54.748.438.315.9Humanity’s Last ExamReasoning with tools62.157.952.251.4OSWorld-VerifiedComputer use80.883.478.776.2SWE-Bench ProReal-repo coding61.569.258.654.2DeepSWE 1.1Long-horizon coding53.359.067.012.0BabyVisionVisual reasoning76.381.283.651.5

Meta-reported, with rivals shown in their strongest modes. Muse Spark 1.1 leads the tool-use and tool-augmented reasoning rows. It places third on coding and multimodal. Consequently, this is an orchestration model, not a coding-accuracy leader. Meta also chose the benchmark set and ran the harness.

The Feature Worth Taking Seriously: Compaction and Delegation

Beyond the scores, orchestration behavior explains the tool-use results. The model actively manages its million-token context window. It remembers actions, retrieves information from much earlier work, and compacts what it keeps.

Delegation is the second half. As a main agent, it gathers context, plans, and delegates execution across parallel subagents. As a subagent, it adheres to its job, understands available tools, and escalates back when needed. The research team also reports zero-shot generalization to new native tools, MCP servers, and custom skills.

Computer use follows the same logic. The model was trained to write scripts when automation is faster. It clicks when direct interaction is simpler. It generates batches of actions at each step.

Wiring It Into an Existing Stack

Because the Model API is OpenAI-compatible, migration is a base-URL change rather than a rewrite. The snippet below is Meta’s own first-call recipe.

pip install openai import os from openai import OpenAI # The OpenAI SDK does not auto-read MODEL_API_KEY, so pass it explicitly. client = OpenAI( base_url="https://api.meta.ai/v1", api_key=os.environ["MODEL_API_KEY"], ) response = client.chat.completions.create( model="muse-spark-1.1", messages=[{"role": "user", "content": "Hello, world!"}], ) print(response.choices[0].message.content)

Anthropic-format harnesses, such as Claude Code, point at the Messages API instead. Agent CLIs like OpenCode register a provider using three values: base URL, key, model ID.

In practice, the shipped demos map onto workloads engineering teams already recognize.

For instance:

In the Facebook Marketplace demo, the model takes smartphone video, extracts useful photos, reasons about the product, then operates a browser to publish the listing.

In an OpenCode demo, it builds a chat web app, takes automated screenshots, traces failures back to code, then validates fixes.

In the dinner-party demo, new context arrives mid-order, and the model updates the plan unprompted.

Coding harnesses get first-class support: planning mode, goal conditioning, subagent delegation, context compaction. Meta team also reports a significant gain on Meta Internal Coding Bench.

Strengths and Weaknesses

Taken together, the trade is legible.

Leads Meta’s reported tool-use and tool-augmented reasoning evaluations

Million-token context, compacted by the model across long sessions

Zero-shot generalization to unfamiliar tools, MCP servers, custom skills

OpenAI and Anthropic SDK compatibility makes an A/B test cheap

Third on SWE-Bench Pro, DeepSWE 1.1, and BabyVision in Meta’s table

Closed weights, so no local deployment and no fine-tuning

Every launch figure is vendor-reported, rivals shown at maximum settings

Public preview reported as US-only; preview pricing can change

Key Takeaways

Tops Meta’s reported tool-use evals; trails Opus 4.8 and GPT-5.5 on coding.

The 1M-token window is actively compacted by the model itself.

Pricing: $1.25 in / $4.25 out per million tokens, plus $20 free credits.

Drop-in for OpenAI and Anthropic SDKs; preview is US-only.

Meta’s first paid API for its own frontier model.