Meta Superintelligence Labs has released Muse Spark 1.1, a new multimodal reasoning model designed for coding, computer use, and complex tasks involving external tools.
Alongside the model, Meta has opened a public preview of the Meta Model API. The launch marks a major shift for a company that previously distributed many of its AI models mainly through downloadable open weights. Developers can now access Muse Spark 1.1 through a hosted, pay-per-token service.
Agentic Capabilities
Meta describes Muse Spark 1.1 as a reasoning model built for agentic tasks. It can plan multi-step work, operate software tools, and complete actions across different applications with limited human involvement.
The model accepts text, images, videos, and PDF documents as input and produces text responses. Meta’s documentation lists a context window of 1,048,576 tokens, allowing it to process large collections of documents or maintain information across lengthy sessions.
Developers can also adjust the amount of reasoning used for each request. The API supports structured outputs, parallel tool calls, file handling, prompt caching, and web searches that return cited answers.
Context Management
One of Muse Spark 1.1’s main features is its ability to manage its own context during long tasks.
The model can remember previous actions, retrieve information from much earlier in a session, and compact its working context while preserving details that may be needed later.
It can also coordinate multiple agents. A main agent can gather information, create a plan, and assign parts of the task to parallel subagents. Those subagents can complete their assigned work and return control when they need additional guidance.
Meta says the model can adapt to unfamiliar native tools, Model Context Protocol servers, and custom skills without requiring additional training.
Computer Use
Muse Spark 1.1 is designed to decide whether a task should be completed through direct interface controls or automated scripts.
For simple actions, it can click, type, and navigate through an interface. For repetitive or complex work, it can write scripts and perform several actions together rather than handling every step separately.
In one demonstration, the model processed a smartphone video, selected suitable product images, identified details about the item, and operated a browser to create a Facebook Marketplace listing.
Another demonstration showed it building a web application, taking screenshots, identifying visible problems, tracing them back to the relevant code, and testing its fixes.
Benchmark Results
Meta’s own testing places Muse Spark 1.1 ahead of Claude Opus 4.8, GPT-5.5 and Gemini 3.1 Pro on several tool-use and tool-assisted reasoning tests.
| Benchmark | Muse Spark 1.1 | Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|---|
| MCP Atlas | 88.1 | 82.2 | 75.3 | 78.2 |
| JobBench | 54.7 | 48.4 | 38.3 | 15.9 |
| Humanity’s Last Exam With Tools | 62.1 | 57.9 | 52.2 | 51.4 |
| OSWorld-Verified | 80.8 | 83.4 | 78.7 | 76.2 |
| SWE-Bench Pro | 61.5 | 69.2 | 58.6 | 54.2 |
| DeepSWE 1.1 | 53.3 | 59.0 | 67.0 | 12.0 |
| BabyVision | 76.3 | 81.2 | 83.6 | 51.5 |
Muse Spark 1.1 led the MCP Atlas, JobBench, and tool-assisted Humanity’s Last Exam results. However, Claude Opus 4.8 performed better in OSWorld-Verified and SWE-Bench Pro, while GPT-5.5 led DeepSWE 1.1 and BabyVision.
The results suggest that Muse Spark 1.1 is strongest as an orchestration and tool-use model rather than as the leading option for coding accuracy or visual reasoning.
However, all figures were published by Meta. The company acknowledged that its testing environments may not have been specifically optimized for competing models, meaning their results may not represent the best performance possible in tools designed around their individual strengths.
Pricing and Availability
Muse Spark 1.1 is available free through the Meta AI app and website when users select Thinking mode.
Developers in the United States can access it through the Meta Model API public preview. Meta is charging $1.25 per million input tokens and $4.25 per million output tokens.
Every new API account receives $20 in free credits for testing prompts and building early integrations. Meta has not yet opened the preview to developers in the European Union.
Developer Integration
The Meta Model API supports the OpenAI SDK and other tools that use OpenAI-compatible request formats.
Developers can connect existing applications through Meta’s API address and use muse-spark-1.1 as the model identifier instead of rebuilding their systems from the beginning.
Meta also supports Anthropic-style message formats for tools and coding environments originally designed around Claude. This makes it easier for developers to compare Muse Spark 1.1 with existing models inside the same workflow.
Strengths and Limitations
Muse Spark 1.1’s main strengths are its tool use, long context window, multimodal inputs, context compaction, and ability to delegate work among multiple agents.
Its relatively low API price and compatibility with existing developer tools also make testing it less expensive and technically easier.
However, it finished behind leading competitors on several coding and visual reasoning benchmarks. The model is hosted by Meta, so developers cannot deploy it locally or fine-tune its weights.
The API also remains in public preview, is initially limited to the United States, and may receive changes to its features or pricing before a wider release.
Stay Connected with ProPakistani
Get the latest tech news, telecom insights, and product launches wherever you prefer.
Add ProPakistani to Preferred Sources and see more of our stories in Google Search and Top Stories.
