Meta AI Releases Muse Code (Beta): A Terminal Coding Agent Powered by the New Muse Spark 1.2 Model
MarkTechPost Asif Razzaq ● Covered by 3 sources
Meta just launched Muse Code, a beta terminal coding agent built on its new Muse Spark 1.2 model. It's hosted-only for now, but it's aimed squarely at big, messy codebases developers actually work in.
Meta wants a piece of the agentic coding wars, and its opening move is Muse Code, a terminal-based agent now in beta alongside a fresh model called Muse Spark 1.2. The pitch isn't just another chatbot that writes snippets. It's a system built to plan, write, and validate changes across sprawling repositories, the kind of work that used to require a human staring at fifteen open files at once.
What sets the runtime apart is persistence. Instead of spinning up a new agent for every task, Muse Code keeps async background agents alive for an entire session, letting them decide on their own when to check back in with the main agent. Meta says this cuts down on redundant digging through the codebase and reduces the babysitting usually required on long, multi-step jobs. Pair that with a local, append-only event log that records every model call, tool run, approval, and edit, and you get something Meta calls replay-exact and restart-safe. Crash mid-task, and the agent picks up exactly where it left off rather than starting over.
The model itself, Muse Spark 1.2, was co-trained directly with the Muse Code harness, which is a detail worth sitting with. Meta fed it rejection-sampled trajectories from actual harness runs and tuned recipes for goal-setting, context compaction, and subagent behavior, meaning the model and the tool were built to fit each other rather than bolted together after the fact. Training also leaned into whole-repository generation and long-horizon projects, with Muse Spark 1.1 generating its own tough coding environments to grade candidate solutions against, a self-improvement loop that fed the 1.2 dataset.
On benchmarks, Meta ran the model through Terminal-Bench 2.1's full 89 tasks, the 113-task DeepSWE v1.1 suite spanning 91 repos, and a 440-task internal bench pulled from real pull requests, all inside isolated Daytona sandboxes. It stacked results against Grok 4.5, Claude Opus 5, GPT-5.6 Terra, Gemini 3.6 Flash, and Kimi K3, though Meta itself flags that its harness probably wasn't tuned for those rivals, so treat the comparisons with some skepticism. Meta also published a case study showing the agent grinding through a GPU kernel optimization task with over 1,000 tool calls across a 24-hour stretch, which is the kind of endurance test that actually says something about reliability.
For now, Muse Code installs via a curl script on macOS and Linux, and Muse Spark 1.2 is reachable through the Meta Model API with wider global access. There's no mention of downloadable weights, so anyone adopting this is signing up for a hosted dependency, not an open model they can run themselves.
My take
Meta calling this a step toward the frontier while keeping the weights closed is the tell here — this is a product play dressed up as research. Co-training a model with its own harness is a smart engineering move and probably does improve real-world reliability, but pairing that with zero downloadable weights means teams adopting Muse Code are locking into Meta's stack the same way they'd lock into any closed vendor. Fine if the tool is genuinely better, but calling it open-ish innovation while it's fully hosted is just marketing.
Read more about this at: MarkTechPost
