Meta Superintelligence Labs Releases Muse Spark 1.1: A Multimodal Reasoning Model for Agentic Tasks on Meta Model API
Today, Meta Superintelligence Labs launched Muse Spark 1.1. Alongside it, Meta opened a public preview of the Meta Model API. That second half is the structural change. Meta’s fashions beforehand reached builders primarily as open weights. Muse Spark 1.1 is closed, hosted, and metered per token. So the query is slender. Where does it belong in a stack you already run?
What is Muse Spark 1.1?
Meta describes it as a multimodal reasoning mannequin constructed for agentic duties. Reported good points over the primary Muse Spark sit in device use, pc use, coding, and multimodal understanding. The context window is 1,000,000 tokens. Meta Model API docs listing 1,048,576.
Core Capabilities and Features
Because it’s a reasoning mannequin, so it thinks earlier than answering. Furthermore, this reasoning effort is adjustable per request. Inputs span textual content, photographs, video, and paperwork; output is textual content. The API additionally exposes structured output, parallel device calling, a Files API, and immediate caching. Adding a web_search device to a Responses API name returns cited solutions.
Pricing and Regional Availability
Access splits two methods. Consumers get it free in ‘Thinking’ mode within the Meta AI app and on meta.ai. Developers pay $1.25 per million enter tokens and $4.25 per million output tokens. New accounts get $20 in free credit. Initial launch publish describes the general public preview as US-only, with no EU entry but.
Performance
With the spec in place, the numbers clarify the positioning. To illustrate this, Meta revealed a launch desk, and the desk splits cleanly.
| Benchmark | Tests | Muse Spark 1.1 | Opus 4.8 (max) | GPT-5.5 (xhigh) | Gemini 3.1 Pro (excessive) |
|---|---|---|---|---|---|
| MCP Atlas | Scaled device use | 88.1 | 82.2 | 75.3 | 78.2 |
| JobBench | Professional device use | 54.7 | 48.4 | 38.3 | 15.9 |
| Humanity’s Last Exam | Reasoning with instruments | 62.1 | 57.9 | 52.2 | 51.4 |
| OSWorld-Verified | Computer use | 80.8 | 83.4 | 78.7 | 76.2 |
| SWE-Bench Pro | Real-repo coding | 61.5 | 69.2 | 58.6 | 54.2 |
| DeepSWE 1.1 | Long-horizon coding | 53.3 | 59.0 | 67.0 | 12.0 |
| BabyVision | Visual reasoning | 76.3 | 81.2 | 83.6 | 51.5 |
Benchmark Analysis
Meta-reported, with rivals proven of their strongest modes. Muse Spark 1.1 leads the tool-use and tool-augmented reasoning rows. It locations third on coding and multimodal. Consequently, that is an orchestration mannequin, not a coding-accuracy chief. Meta additionally selected the benchmark set and ran the harness.
The Feature Worth Taking Seriously: Compaction and Delegation
Beyond the scores, orchestration conduct explains the tool-use outcomes. The mannequin actively manages its million-token context window. It remembers actions, retrieves info from a lot earlier work, and compacts what it retains.
Delegation is the second half. As a major agent, it gathers context, plans, and delegates execution throughout parallel subagents. As a subagent, it adheres to its job, understands out there instruments, and escalates again when wanted. The analysis crew additionally stories zero-shot generalization to new native instruments, MCP servers, and customized expertise.
Computer use follows the identical logic. The mannequin was educated to write down scripts when automation is quicker. It clicks when direct interplay is less complicated. It generates batches of actions at every step.
Wiring It Into an Existing Stack
Because the Model API is OpenAI-compatible, migration is a base-URL change reasonably than a rewrite. The snippet under is Meta’s personal first-call recipe.
# pip set up openai
import os
from openai import OpenAI
# The OpenAI SDK doesn't auto-read MODEL_API_KEY, so cross it explicitly.
shopper = OpenAI(
base_url="https://api.meta.ai/v1",
api_key=os.environ["MODEL_API_KEY"],
)
response = shopper.chat.completions.create(
mannequin="muse-spark-1.1",
messages=[{"role": "user", "content": "Hello, world!"}],
)
print(response.selections[0].message.content material)Anthropic-format harnesses, reminiscent of Claude Code, level on the Messages API as a substitute. Agent CLIs like OpenCode register a supplier utilizing three values: base URL, key, mannequin ID.
Use Cases
In observe, the shipped demos map onto workloads engineering groups already acknowledge.
For occasion:
Multimodal itemizing automation:
- In the Facebook Marketplace demo, the mannequin takes smartphone video, extracts helpful photographs, causes concerning the product, then operates a browser to publish the itemizing.
Screenshot-driven debugging:
- In an OpenCode demo, it builds a chat net app, takes automated screenshots, traces failures again to code, then validates fixes.
Adaptive planning:
- In the dinner-party demo, new context arrives mid-order, and the mannequin updates the plan unprompted.
Coding harnesses get first-class assist: planning mode, objective conditioning, subagent delegation, context compaction. Meta crew additionally stories a major acquire on Meta Internal Coding Bench.