Immediately, we’re excited to introduce Muse Spark 1.1, the most recent mannequin from Meta Superintelligence Labs and a major improve from Muse Spark. Muse Spark 1.1 is a multimodal reasoning mannequin constructed for agentic duties, with main beneficial properties in instrument and laptop use, coding, and multimodal understanding.
With these enhancements, Muse Spark 1.1 advances the performance-efficiency frontier. Along with this week’s launch of Muse Image, this launch brings us nearer to our imaginative and prescient of private superintelligence: fashions that aid you pursue your objectives, create what you think about, deepen your relationships, and take motion on what you worth most.
Together with this launch, we’re launching a public preview of the brand new Meta Model API the place builders can entry Muse Spark 1.1. The mannequin is on the market now in “Considering” mode within the Meta AI app and on meta.ai.
Evaluations
Brokers
Muse Spark 1.1 delivers distinctive efficiency in private agentic duties that require planning and orchestration throughout a spread of exterior apps and providers. It zero-shot generalizes to new native instruments, MCP servers, and customized abilities.
It tackles advanced initiatives considerably quicker than Muse Spark, as it’s educated to orchestrate multi-agent methods to optimize end-to-end latency. As the principle agent, it could collect context, make a plan, and delegate execution throughout parallel subagents. As a subagent, it adheres to its job, understands accessible instruments, and is aware of when to escalate again to the principle agent.
Muse Spark 1.1 can actively handle its context window of 1 million tokens. It remembers actions, retrieves info from a lot earlier work, and compacts in a approach that retains the vital steps wanted for later work.
Muse Spark 1.1 excels at computer-use workflows that unfold throughout a number of purposes with info altering on-the-fly. It maintains context throughout prolonged classes, adapts to evolving necessities, and navigates unfamiliar interfaces with minimal human intervention.
Somewhat than reasoning by means of each desktop the 1st step click on at a time, Muse Spark 1.1 understands when to automate and when to make use of the interface straight. We educated the mannequin to jot down scripts when automation is quicker, click on when direct interplay is less complicated, and generate batches of actions at every step.
Agentic feast group: In real-world purposes, new context arises that adjustments the duty. Muse Spark 1.1 notices these adjustments when inserting the dinner order and makes crucial updates with out consumer intervention.
Coding
Coding efficiency for Muse Spark 1.1 improved considerably on real-world duties involving giant, advanced codebases. It will possibly diagnose and repair advanced bugs, implement new options in enterprise-grade methods, and execute giant code migrations. In use instances like creating internet purposes and end-to-end query answering, Muse Spark 1.1 exhibits giant beneficial properties over our first mannequin.
We educated our mannequin to easily adapt to numerous harnesses and reliably deal with advanced multi-turn dynamics. Muse Spark 1.1 performs effectively with common agentic coding setups, supporting widespread options like planning mode, purpose conditioning, subagent delegation, and context compaction.
Debugging demo in OpenCode: Muse Spark 1.1 builds a chat internet app, takes automated screenshots to establish user-visible failures, traces points again to related code to implement fixes, and validates these adjustments. The mannequin seamlessly combines coding, multimodal understanding, and power calling.
Throughout Meta, builders and researchers are utilizing Muse Spark 1.1 each day to construct quicker and work smarter. On our main inside coding analysis, Meta Inside Coding Bench, Muse Spark 1.1 considerably improves upon Muse Spark and is aggressive with main alternate options.
Researchers are actually additionally automating mannequin improvement and analysis duties by leveraging Muse Spark 1.1 of their workflows.
DeepSWE analysis in OpenCode: Muse Spark 1.1 evaluates itself on a subset of DeepSWE duties throughout completely different reasoning strengths and produces an evaluation dashboard primarily based on the outcomes.
Together with coding and agentic capabilities, Muse Spark 1.1 excels in notion, multimodal reasoning, and power use. It will possibly work together with actual environments and produce grounded outputs with strengths in visual-to-code artifact technology, ultra-descriptive picture and video captioning, and agentic workflow execution for multimodal use instances.
Muse Spark 1.1’s multimodal capabilities are particularly precious when notion and motion must occur collectively. The mannequin can examine visible and audio, protect particulars throughout an extended workflow, and use these particulars whereas working computer systems on the consumer’s behalf.
Fb Market agent: Utilizing video shot from a smartphone, Muse Spark 1.1 extracts helpful photographs and causes concerning the product to function a consumer’s browser and make a Fb Market itemizing on the consumer’s behalf.
Security
We performed intensive security evaluations earlier than deployment, following the Advanced AI Scaling Framework, which defines evaluations, menace fashions, and deployment thresholds for our most superior fashions.
Throughout all frontier danger classes — Chemical & Organic, Cybersecurity, and Lack of Management — our evaluations present Muse Spark 1.1 operates inside secure margins. Muse Spark 1.1 demonstrates robust resistance to direct jailbreaks and oblique assaults from untrusted knowledge, immediate injection, and developer-prompt assaults. Consequently, it exhibits higher adversarial robustness, decrease hallucination charges, and decreased sycophancy.
Our full security posture for 1.1 is documented in our Muse Spark 1.1 Evaluation Report.
Availability
For the primary time, builders can start constructing with Muse Spark 1.1 by way of the brand new Meta Mannequin API, now in public preview. Early companions of Muse Spark 1.1 reward the mannequin as a whole agentic basis, pairing lengthy context dealing with with robust coding and reasoning capabilities to deal with large-scale agentic workloads.
“What’s most spectacular about Muse Spark is how a lot it packs into one mannequin: large million-token context, full multimodal help (photos, video, PDFs), built-in search with citations, robust reasoning, top-tier coding talents (notably frontend and design), structured output, and parallel instrument calling — all in a clear OpenAI-compatible bundle. A whole agentic basis.”
— Amjad Masad, CEO of Replit
“Meta is clearly constructing for severe agentic coding – robust instrument use at a value level that makes it viable to run actual coding workloads at scale. That mixture is uncommon, and it’s precisely why we needed Cline builders to have entry early.”
— Saoud Rizwan, CEO of Cline
“When examined towards Field’s enterprise work analysis set, Muse Spark delivered enterprise capabilities aggressive with at the moment’s main frontier fashions. That degree of intelligence, mixed with its strengths in structured, procedural workflows throughout industries reminiscent of skilled providers, public sector, and industrial operations, makes it a compelling alternative for organizations.”
— Yashodha Bhavnani, VP of AI Merchandise at Field
We’re thrilled to be releasing Muse Spark 1.1, a testomony to our analysis momentum. We’ve got much more succesful fashions in coaching and stay up for sharing what’s to return.
Written by:
Meta Superintelligence Labs