Muse Spark 1.1

NanoGPT published independent benchmark analysis on July 18, 2026 showing Meta's Muse Spark 1.1 scored 51 on the Artificial Analysis Intelligence Index, up from 43 for the original Muse Spark released three months earlier. The piece situates that score near GPT-5.4 at xhigh reasoning, GPT-5.6 Luna at max, and GLM-5.2 a The article frames Meta's first-party launch numbers as distinct from independent testing, noting Muse Spark 1.1 is text-and-image capable, supports tool calling, offers adjustable reasoning effort, and accepts up to one million input tokens. According to NanoGPT, the model's standout practical combination is strong be

Meta Superintelligence Labs released Muse Spark 1.1 on July 9, 2026, positioning it as a multimodal reasoning model optimized for agentic tasks, tool use, computer interaction, and coding, with availability through the new Meta Model API and in Meta AI. The model features a 1 million token context window for long-horiz Reported benchmark highlights include MCP Atlas (scaled tool use) at 88.1, JobBench at 54.7 (up from 17.0 for prior Muse Spark, beating Opus 4.8's 48.4 and GPT 5.5's 38.3), and OSWorld-Verified at 80.8, with the article also referencing Humanity's Last Exam. The model supports MCP for tool use, custom skills, structure

Meta released Muse Spark 1.1 on July 9, 2026, alongside a public preview of the Meta Model API, positioning it as a multimodal reasoning model built for agentic tasks with gains in tool use, computer use, coding, and multimodal understanding. The model is described as a significant upgrade from Muse Spark and updates t Rather than booting a pure benchmark leader, Muse Spark 1.1 is positioned as an agent foundation for AI founders, developer-tool companies, and automation startups targeting agents, tools, coding loops, browser work, multimodal workflows, and long-running tasks. The key framing is whether it can complete useful work at

Meta Superintelligence Labs released Muse Spark 1.1 on July 9, 2026 as a multimodal reasoning model built for agentic tasks with major gains in tool and computer use, coding, and multimodal understanding. The model manages a 1-million-token context window and orchestrates multi-agent systems by acting as a main agent that delegates to parallel subagents or as a subagent that escalates back. A public preview of the Meta Model API launched alongside, and the model is available in "Thinking" mode on the Meta AI app and meta.ai. For computer use, the model navigates unfamiliar interfaces across extended sessions, deciding when to write automation scripts versus click directly and generating batches of actions per step. It zero-shot generalizes to new native tools, MCP servers, and custom skills while remaining competitive against frontier alternatives. The release accompanied the launch of Muse Image and advances Meta's vision of personal superintelligence for goal pursuit, creativity, and action.

The BenchLM tracker for Muse Spark 1.1 reports a 65.9/100 capability score, with verified category rankings of 22nd in Agentic (57.7, 80th percentile across 14 benchmarks), 23rd in Coding (56.3, 84th percentile across 4 benchmarks), 15th in Knowledge (68.2, 91st percentile across 5 benchmarks), and non-ranked but verif The page flags Muse Spark 1.1 as "Superseded" by Muse Spark 1.2, with the data snapshot dated September 24, 2026, indicating the variant has aged relative to newer Muse Spark releases. Despite the supersession status, the tracker still offers direct benchmark comparisons against the broader ranked and priced model cata

Meta released Muse Spark 1.1 on July 9, 2026 as a proprietary multimodal reasoning model with a 1-million-token context window, 256k max output, and API pricing of $1.25 input and $4.25 output per million tokens. The model accepts text, image, video, and file inputs and supports reasoning, with Vals.ai tracking its scores across 15 benchmarks including MedScribe (88.89%), GPQA Diamond (91.16%), TaxEval v2 (79.72%), and Harvey's Legal Agent Benchmark (20%). On agentic benchmarks, Muse Spark 1.1 scored 57.21% on Finance Agent v2 and 45.58% on SAGE, with 31.11% on Code Migration and 72.16% on Vibe Code Bench v1.1. The model was later superseded in Meta's lineup by Muse Spark 1.3 Max on September 3, 2026 and Muse Spark 1.3 on September 1, 2026, though the 1.1 version remains a tracked benchmark baseline. Average measured latency on Vals.ai was 22 minutes 11 seconds at an accuracy of 60.28%.

Meta Superintelligence Labs announced Muse Spark 1.1 on July 9, 2026, a multimodal reasoning model built for autonomous agentic workflows with a 1-million-token context window and upgraded coding and computer control capabilities. The company simultaneously launched a public preview of the Meta Model API, opening the Muse architecture to outside developers for the first time. The model is live in "Thinking" mode on the Meta AI app and meta.ai. The model can act as both a "primary agent" and a "subagent," breaking large tasks into parallel delegations to cut processing latency, and can automate multi-application workflows through scripts or direct interface interaction. Early API partners Replit, Cline, and Box headlined the announcement, with Replit CEO Amjad Masad citing the million-token context plus multimodal support as transformative. Internal Meta testing on the Meta Internal Coding Bench reportedly shows Muse Spark 1.1 outperforming its predecessor competitively against unnamed industry alternatives.

Meta Superintelligence Labs released Muse Spark 1.1 on July 9, 2026 as a multimodal reasoning model targeting agentic workflows, marking a major upgrade over the original Muse Spark with gains in tool and computer use, coding, and multimodal understanding. The model actively manages a 1-million-token context window and is trained to orchestrate multi-agent systems as either a main agent or subagent. Along with the release, Meta launched a public preview of the Meta Model API for developer access, and Muse Spark 1.1 is available in "Thinking" mode on the Meta AI app and meta.ai. For computer use, Muse Spark 1.1 is trained to write automation scripts when faster than direct interaction and to click when simpler, generating batches of actions across multiple applications with minimal human intervention. It zero-shot generalizes to new native tools, MCP servers, and custom skills, and can compact context while preserving critical steps for later work. The launch was paired with Muse Image, advancing Meta's stated vision of personal superintelligence.