Muse Spark 1.2 (Meta)

Labellerr's third-party blog, dated 2026-08-11, reports that Meta released Muse Spark 1.2 on August 5, 2026 as the direct successor to Muse Spark 1.1, positioning it as a coding-focused language model that plans changes, writes code, and validates results across large repositories. The article states Meta scaled up tra The piece frames Muse Spark 1.2 as an end-to-end software engineering tool that coordinates multiple tasks at once when paired with Muse Code, handling planning, code generation, and self-checking on multi-file codebases. Meta's broader goal, as relayed by the blog, is to build models that act like real human developer

On 5 August 2026, Meta launched Muse Code into public beta, a dedicated AI coding agent built on a new model called Muse Spark 1.2. According to the Tech Insider Ireland recap of the launch, Muse Spark 1.2 was co-trained specifically for the agent rather than adapted from a general-purpose chat model, marking a deliber The same recap describes Muse Code as a terminal-based agent that installs with a single command, plans, writes, and validates code across large repositories, and splits large jobs across parallel sub-agents. It also outlines a two-tier pricing structure that trades cheaper tokens for allowing Meta to train on user pro

The BenchLM tracker page for Muse Spark 1.2 catalogues the model's published numbers: capability 70.4/100, 1M-token context window, ~154 tokens/sec output speed, and an API price of $1.25 per million input tokens and $4.25 per million output tokens (with a $0.15 cached-input rate). The page marks Muse Spark 1.2 as supe Across BenchLM's category percentile breakdown, Muse Spark 1.2 ranks at the 95th percentile for knowledge (rank 10 of 183), 89th percentile for agentic work, and 87th percentile for coding, while reasoning, math, multilingual, multimodal, and instruction-following axes are listed as not eligible. Coding is flagged as t