Meta's Muse Code: A Coding Agent — But Is the Model Open?

On August 5, 2026, Meta launched Muse Code (beta), a terminal-based coding agent powered by the new Muse Spark 1.2 model. For developers who've built their world on Meta's open Llama models, the launch raises one central question: Is the model open? And if not — what's the plan?

The short answer is no — but with an important caveat about the future.

What is Muse Code?

Muse Code is a coding agent you run directly in your terminal:

curl -fsSL https://dev.meta.ai/install.sh | bash

It handles complex software engineering tasks across large repositories — planning changes, writing code, and validating the results itself.

Three technical highlights stand out:

Async background agents. Instead of spawning helper agents per task, Muse Code keeps a set of specialized background agents active throughout the entire session. They carry out the next steps on their own and decide for themselves when to report back to the main agent — reducing latency on hard, multi-step tasks.

Crash-safe event log. Everything — every model call, tool execution, approval, and edit — is logged to a local event log. That makes the runtime replay-exact and restart-safe: after a crash, the agent can resume exactly where it left off. That makes it possible to take on long-running tasks without being derailed by errors.

Built-in skills. /plan turns a task into an approval-gated plan; /grill stress-tests the plan until it holds up; /goal works toward completing the goal.

What is Muse Spark 1.2?

Muse Spark 1.2 is a coding-focused update to Muse Spark 1.1 (from July) with improvements in code generation, complex debugging, codebase understanding, and end-to-end developer workflows.

The model was co-trained with the Muse Code agent itself to ensure the best possible performance when used together — including rejection-sampled harness trajectories and optimization for goals, compaction, and subagents.

Two notable aspects:

Long-horizon training. The model was trained intensively on long-duration coding tasks — whole-repository generation, large end-to-end projects, and auto-research — using planning to sequence work and context compaction to retain knowledge over time.

Self-improvement loop. Meta used Muse Spark 1.1 to generate challenging coding environments; the model then graded its own solutions against the requirements — a scalable training dataset that made the successor better at complex instructions than its predecessor.

In one case study, Meta had the agent iteratively optimize GPU kernels over more than 1,000 tool calls (up to 24 hours) on NVIDIA Hopper GPUs, with significant improvements over the baseline implementation.

Is the model open? No — but the plans are interesting

Here's the central question for this blog's readers: Is Muse Spark open?

No. The Muse Spark family has closed weights. The model is only available via the Meta Model API, not as downloadable weights:

• There is no repository for either Muse Code or Muse Spark in Meta's meta-llama or facebook Hugging Face organizations.
• Parameter count and architecture are not disclosed.
• The license "doesn't make sense" precisely because the weights are closed — API use is governed by preview terms.
• Even the installer downloads a proprietary binary via a shell script; there's no public source repository for the agent itself at launch.

That marks a significant break from the Llama lineage that made Meta AI's open champion for three years and over a billion downloads.

But what does Meta say about the future?

Meta has repeatedly signaled that the closure is strategic rather than permanent:

"Meta says it 'hopes to open-source future versions' of the Muse series itself." — TNW

And according to reports from Axios/SiliconANGLE, Meta is actively working on open-weight versions of its upcoming AI models, as a complement to the proprietary flagships:

"Open-weight versions of these models will follow." — NYU RITS

There are, however, clear caveats:

• Future open-weight models will likely not include all the capabilities of their proprietary counterparts.
• The largest frontier models may well remain closed.
• Meta specifically plans to keep certain capabilities proprietary — particularly around cybersecurity code generation and certain mixture-of-experts components.

This paints a picture of a hybrid strategy under Alexandr Wang's Superintelligence Labs: proprietary models for Meta's own consumer products (Meta AI in WhatsApp, Instagram, etc.), open-weight releases for the developer ecosystem — with Wang himself describing Meta as "a counterweight to Anthropic and OpenAI."

The pricing tells its own story

The pricing page underscores just how far we are from the old Llama model:

The Contributor tier gives a ~12–21x discount — but only if you grant Meta the right to train future models on your prompts and completions (with a rate limit of just 60 req/min versus 3,000 for standard). It's really a data-governance choice more than a pricing question: the discount only exists where Meta is allowed to train on your traffic.