ainewsblitz.com

Breaking

Meta's Muse Spark 1.2 Tops Video-to-Website Ranking at 1279 Elo

  • Foundation Models
  • AI Agents
  • Software Dev & Coding

Meta's Muse Spark 1.2 has taken first place in Design Arena's Video-to-Website leaderboard with an Elo rating of 1279, while ranking second in Image-to-HTML at 1252 and third in Image-to-Frontend. The results strengthen the model's profile in visual software generation, an area Meta highlighted in its recent multimodal demonstrations. Meta introduced Muse Spark 1.2 on August 5 alongside Muse Code, a beta terminal coding agent designed to plan, edit and validate changes across large repositories. Muse Code is available for macOS and Linux, while Muse Spark 1.2 is offered through Muse Code and the Meta Model API with expanded global access. Meta's announcement describes persistent background agents, a restart-safe local event log and built-in skills for planning, reviewing and pursuing long-running objectives. In a separate August 20 update, Meta said Muse Spark 1.2 can convert images and videos into working web pages and games, then inspect the generated artifacts and refine them through a self-improvement loop. Meta said the Design Arena rankings reflected model inclusion and leaderboard positions as of August 18. The multimodal update also framed the model's strongest gains as tool-assisted visual reasoning rather than passive image understanding. The two listed Elo scores are 27 points apart, but they measure different input tasks; the result should therefore be read as evidence of strength in visual web construction, not as a universal model ranking. On the engineering side, Meta says Muse Spark 1.2 was trained for code generation, complex debugging, codebase understanding and long-horizon work. Its published case study reports iterative GPU-kernel optimization over more than 1,000 tool calls and up to 24 hours on NVIDIA Hopper hardware. A current model profile lists API pricing at $1.25 per million input tokens, $4.25 per million output tokens and $0.15 per million cached input tokens, with a context window close to one million tokens and a maximum generation length of 131,072 tokens. Roboflow's model page reports an 80.4% average across its six-task Vision Evals suite, with particularly strong OCR performance but weaker relative results in object detection and data extraction. Early hands-on reception is mixed in a useful way. A review of the community KingBench 3 suite gave Muse Spark 1.2 a reported 61/80, or 76.25%, with strong results on visual frontend work, Three.js and SVG generation, while game mechanics and detailed state management were less dependable. That review cautioned that KingBench is a community stress test rather than a standardized industry evaluation. Another independent test built five applications and reported that the model repeatedly struggled to enforce logic rules it appeared to understand, including traffic-light behavior and game win conditions, even while producing usable frontend projects. The hands-on account also estimated roughly $10 in Standard-tier usage for the five builds. The emerging picture is of a model that can translate rich visual instructions into attractive, interactive interfaces and sustain ambitious agentic sessions, but still requires verification when correctness depends on hidden state, backend logic or edge cases. Design Arena's latest placements make visual coding a clear area to watch; broader adoption will depend on whether those frontend strengths carry over to production repositories and less forgiving software workflows.