Muse Spark 1.1 Release Notes

Muse Spark 1.1: A Multimodal Foundation for Agentic AI

Meta Superintelligence Labs has introduced Muse Spark 1.1, a significant upgrade to the Muse Spark model designed specifically for agentic tasks. The model focuses on improving tool use, computer interaction, coding, and multimodal understanding to enable more autonomous AI agents.

Agentic Orchestration and Planning

Muse Spark 1.1 is engineered to handle complex, multi-step projects by orchestrating multi-agent systems to reduce end-to-end latency. The model can act as a primary agent that gathers context and delegates tasks to parallel subagents, or as a subagent that executes specific tools and escalates issues back to the main agent when necessary.

Key agentic capabilities include:

  • Zero-shot Generalization: The model generalizes to new native tools, MCP servers, and custom skills without prior training.
  • Context Management: It features a 1-million-token context window, allowing it to remember previous actions and retrieve information from early stages of a workflow while compacting data to maintain critical steps.

Computer Use and Automation

Muse Spark 1.1 is optimized for workflows that span multiple applications where information changes dynamically. Rather than relying solely on single-click interactions, the model determines the most efficient path to completion:

  • Scripting: The model writes scripts when automation is faster than manual interaction.
  • Direct Interaction: It uses the interface directly (clicking) when that is simpler.
  • Batching: It generates batches of actions at each step to increase efficiency.

Enterprise-Grade Coding and Debugging

Coding performance has been improved for large, complex codebases. Muse Spark 1.1 can diagnose bugs, implement new features in enterprise systems, and execute large-scale code migrations.

In the OpenCode environment, the model demonstrates a combined workflow of coding, multimodal understanding, and tool calling by taking automated screenshots of user-visible failures, tracing them back to the source code, and implementing fixes.

Multimodal Perception and Action

The model integrates perception and action, allowing it to operate computers on a user's behalf based on visual and audio inputs. For example, in a Facebook Marketplace use case, Muse Spark 1.1 can analyze a smartphone video of a product, extract relevant photos, and reason about the product to create a listing in a web browser.

Safety and Robustness

Meta developed Muse Spark 1.1 following the Advanced AI Scaling Framework. Evaluations across frontier risk categories—including Cybersecurity, Chemical & Biological, and Loss of Control—indicate the model operates within safe margins. The model shows increased resistance to prompt injection, direct jailbreaks, and indirect attacks from untrusted data, resulting in lower hallucination rates and reduced sycophancy.

Availability and Developer Access

Muse Spark 1.1 is available in "Thinking" mode within the Meta AI app and on meta.ai. For developers, Meta has launched a public preview of the Meta Model API, providing an OpenAI-compatible package for building agentic applications.

Industry Feedback

  • Replit CEO Amjad Masad described the model as a "complete agentic foundation," citing its million-token context, multimodal support, and parallel tool calling.
  • Cline CEO Saoud Rizwan highlighted the model's strong tool use and price point as viable for running real coding workloads at scale.
  • Box VP of AI Products Yashodha Bhavnani noted that the model's intelligence is competitive with leading frontier models for enterprise work evaluations.

Community Discussion and Critiques

While the release has been met with interest regarding its capabilities and pricing, several technical and ethical concerns were raised by the community:

  • Benchmarking Integrity: Some users questioned the validity of the model's benchmarks. One former Meta employee noted that the Terminal-bench-2.1 results appeared to override CPU and RAM limits, which would typically result in disqualification from official leaderboards.

  • Data Privacy: Multiple users expressed concern over the lack of a clearly stated data retention policy for the paid API, questioning how user data is handled compared to other industry providers.

  • Data Accessibility: Users reported regional restrictions for the API, with some noting that the Model API was unavailable in their specific regions despite not being on a restricted list.

  • Model Nature: There was disappointment among some developers that Muse Spark 1.1 is a closed-weights model, contrasting with Meta's previous history of releasing open-weights models.

Sources

Related