Meta launched Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, its up to date coding mannequin. It is obtainable now by way of the Meta Mannequin API and a curl set up script.
The agent coordinates persistent background subagents and retains a replay-exact occasion log, so a crash resumes precisely the place it stopped.
On Meta’s personal charts, Muse Spark 1.2 trails Anthropic’s Opus 5 on each coding benchmark proven, whereas beating OpenAI’s Codex and Google’s Antigravity on most.
Meta is the most recent tech big to ship a coding agent, racing to compete with main AI behemoths Anthropic and OpenAI.
“We’re excited to launch Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, our latest mannequin,” the corporate wrote in an official announcement. “This marks our subsequent step towards the frontier, with bigger and rather more succesful fashions on the best way.”
As an agentic coding device, Muse Code is constructed for software program engineering throughout giant repositories. Per Meta, it “takes on complicated software program engineering duties throughout giant repositories: planning adjustments, writing code, and validating the outcomes. It will probably coordinate a number of persistent subagents for every activity, fixing troublesome issues quicker, extra precisely, and with much less intervention.”
The element that stands out is the runtime. Muse Code logs each mannequin name, device run, approval, and edit to a neighborhood occasion log that acts as a single supply of fact. “This single supply of fact makes the runtime replay-exact and restart-safe: after a crash, the agent can resume exactly the place it stopped,” Meta stated. For long-running jobs, that is the function that issues greater than uncooked pace—and it is the half rivals have not made a promoting level.
It additionally ships with default abilities. The “/plan” command turns a activity into an approval-gated plan, whereas “/grill” stress-tests that plan till it holds up and “/purpose” works towards profitable completion of the target much like what Hermes does. Meta stated it co-trained Muse Spark 1.2 with Muse Code so the core LLM and the agent work collectively in synergy.
The benchmarks, and the catch
Muse Spark 1.2 is a coding-focused replace to Muse Spark 1.1. Meta stated it “considerably scaled up coaching compute on coding duties whereas increasing coaching setting range, delivering enhancements in code era, complicated debugging, and end-to-end developer workflows.” The charts inform a transparent story.
On Terminal-Bench 2.1, Muse Spark 1.2 with Muse Code scored 82.9%, behind Claude Code on Opus 5 at 86.7% however forward of GPT-5.6 Terra on Codex (81.8%) and Grok Construct (81.6%).
DeepSWE 1.1, which measures agentic coding capabilities, was nearer: 59.3% for Muse versus 65.0% for Opus 5 and 64.8% for Codex. On Meta’s inner coding bench, Muse hit 70.6% to Opus 5’s 79.4%.
The speedup charts flip the order. Over 1,000-plus device calls, Opus 5 posted the largest achieve versus baseline (about 74–75%), with Muse Spark 1.2 mid-pack at roughly 61–69% relying on the run. Meta’s level is that the agent retains bettering as device calls accumulate, the conduct you need from a long-horizon coder.
Probably the most fascinating demos are long-horizon and multimodal. In stress testing, Meta stated Muse Code “iteratively optimized GPU kernels over 1,000+ device calls (as much as 24 hours) on Nvidia Hopper GPUs.” Meaning it was in a position to enhance over time.
There’s additionally a visual-coding angle. In a single demo, a person drops a fly-through video of a home into the terminal as an mp4, and Muse Code “interprets the video and produces a visually wealthy web site with reserving capabilities.” Studying uncooked video right into a working net app is the multimodal pitch Meta has been making throughout the Muse line.
See the launch thread:
Right here’s an instance of Muse Code’s multimodal visible coding capabilities. On this demo, the person inputs a fly-through video of a house into the terminal as an mp4 file. Muse Code interprets the video and produces a visually wealthy web site with reserving capabilities. pic.twitter.com/3CAfIMYmAB
— AI at Meta (@AIatMeta) August 5, 2026
The sphere is already crowded
That stated, Meta is late to the battle. OpenAI’s Codex already runs parallel cloud brokers; DeepSeek has constructed its personal rival to Claude Code and agentic instruments like Hermes or OpenClaw are already good substitutes with extra capabilities. Muse Code’s edge is the crash-safe runtime and the subagent design, not benchmark supremacy.
The danger is the same old one for agentic coding: an agent that resumes after a crash and retains calling instruments for twenty-four hours is highly effective and unpredictable. Meta is betting builders need that autonomy, and it is transport now.
Muse Code is out there for testing upon set up coming into this command:curl -fsSL https://dev.meta.ai/set up.sh | bash
Each day Debrief Publication
Begin daily with the highest information tales proper now, plus authentic options, a podcast, movies and extra.
