Introducing Muse Code (beta), a terminal coding agent built for long-horizon software engineering, powered by our new Muse Spark 1.2 model. Muse Code plans, implements, and validates complex, multi-file changes across large repositories with persistent sub-agents that solve difficult problems faster, more accurately, and with less intervention. 🧵👇

Aug 5, 2026 · 7:25 PM UTC

199
354
3,683
517,956
Under the hood, Muse Code pairs a simple agent loop with persistent, async background agents that stay active throughout each session to execute next steps, avoid redundant info gathering, and reduce steering on complex tasks. An append-only local event log tracks every model call, tool run, approval, and edit, making the runtime replay-exact and restart-safe for long-running workflows.
3
6
162
21,391
Muse Spark 1.2 has improved coding capabilities compared to its predecessor. We significantly scaled up training compute on coding tasks while expanding training environment diversity, delivering improvements in code generation, complex debugging, and end-to-end developer workflows. The model was co-trained with Muse Code to optimize performance when paired together, and focused on tasks like whole-repo generation, large projects, and auto-research.
1
11
189
56,749
In long-horizon stress testing, Muse Code iteratively optimized GPU kernels over 1,000+ tool calls (up to 24 hours) on @nvidia Hopper GPUs, delivering very competitive performance gains for KDA and MLA relative to baseline Triton implementations.
The baseline is the FLA Triton implementation of KDA. Models were prohibited from importing third-party kernel libraries such as FLA directly; instead, they had to apply specialized kernel-optimization knowledge to implement the algorithm in Triton, rather than wrap existing implementations. Muse Spark 1.2 paired a chunk-parallel preparation kernel with a sequential inter-chunk scan, combining standard fusion and tiling with KDA-specific optimizations such as re-centering the gated cumulative decay at the chunk midpoint.

ALT The baseline is the FLA Triton implementation of KDA. Models were prohibited from importing third-party kernel libraries such as FLA directly; instead, they had to apply specialized kernel-optimization knowledge to implement the algorithm in Triton, rather than wrap existing implementations. Muse Spark 1.2 paired a chunk-parallel preparation kernel with a sequential inter-chunk scan, combining standard fusion and tiling with KDA-specific optimizations such as re-centering the gated cumulative decay at the chunk midpoint.

We benchmark against a PyTorch reference implementation at batch size 1, number of heads 64, sequence length 8192, and latent dimension 512. Muse Spark 1.2 designed a two-kernel Triton pipeline for this workload, combining kernel fusion and tiling with MLA-specific optimizations such as reusing the shared KV latent as both K and V.

ALT We benchmark against a PyTorch reference implementation at batch size 1, number of heads 64, sequence length 8192, and latent dimension 512. Muse Spark 1.2 designed a two-kernel Triton pipeline for this workload, combining kernel fusion and tiling with MLA-specific optimizations such as reusing the shared KV latent as both K and V.

1
5
121
14,223
Here’s an example of Muse Code’s multimodal visual coding capabilities. In this demo, the user inputs a fly-through video of a home into the terminal as an mp4 file. Muse Code interprets the video and produces a visually rich website with booking capabilities.
3
22
234
26,136
Muse Spark 1.2 is available today in Muse Code and in Meta Model API, with expanded global access and many new features on the horizon. ➡️ Learn more: go.meta.me/musecode

Introducing Muse Code and Muse Spark 1.2

Introducing Muse Code, a terminal coding agent powered by Muse Spark 1.2, with persistent background agents, repository-scale execution, and built-in verification.

research.meta.ai
2
11
155
20,672
Sort replies: Relevant Recent Liked
Replying to @AIatMeta
Where do we find this?
1
12
7,497
Try it out here: dev.meta.ai

Meta Model API

Meta Model API Developer Portal

dev.meta.ai
2
39
5,761
Replying to @AIatMeta
Meta two tier pricing is smart and refreshingly transparent. For many POC s of mine and non-sensitive workloads I m fine with Meta using the data to improve its product and the contributor pricing is almost DeepSeek-level. Even the second tier is priced better than many competitors. This feels like the beginning of the AI price war.
4
2
84
7,348
It's time to build 🚀
4
43
6,603
Replying to @AIatMeta
Wen Windows?
15
2,016
Replying to @AIatMeta
Great release
Vals AI
@ValsAI
Aug 6
Muse Spark 1.2 just cracked the top 5 on the Vals Index, at just $0.69 per test. This is 3x cheaper than Kimi and 10x or more cheaper than Fable, Opus, and 5.6 Sol.
7
458
Replying to @AIatMeta
Congratulations 🔥, hopefully this speed continues, and September seems to be a good month for SOTA image and video models😉
6
531
Replying to @AIatMeta
▲ Marktechpost AI
@Marktechpost
Aug 5
Meta AI released Muse Code (in beta mode), a terminal coding agent in beta, powered by its new Muse Spark 1.2 model. Here are some key takeaways: 1. Async background agents that outlive the task Muse Code runs a simple agent loop plus a set of specialized background agents. These stay active throughout the session instead of being spawned per task. → Meta says this avoids redundant information gathering → The agents carry out next steps and choose when to report back to the main agent → Stated effect: lower latency and less steering on multi-step tasks 2. An append-only event log as the single source of truth Every model call, tool run, approval, and edit is appended to a local event log. → Meta calls the runtime replay-exact and restart-safe → After a crash, the agent resumes precisely where it stopped → This is what makes long-running tasks survive failures 3. Three bundled skills → /plan turns a task into an approval-gated plan → /grill stress-tests that plan until it holds up → /goal works toward completion of the specified objective 4. Muse Spark 1.2 was co-trained with the harness Training included rejection-sampled harness trajectories and recipe optimizations for goals, compaction, and subagents. Meta also integrated the Muse Code toolset directly to maximize harness compatibility. Long-horizon training covered whole-repository generation, large end-to-end projects, and auto-research. 5. The kernel optimization case study Meta ran iterative GPU kernel optimization over 1,000+ tool calls, up to 24 hours per run. → Benchmarked on KDA and MLA kernels for NVIDIA Hopper GPUs → KDA baseline is the FLA Triton implementation, with third-party kernel libraries prohibited → MLA reference is PyTorch at batch size 1, 64 heads, sequence length 8192, latent dimension 512 → For MLA, the model built a two-kernel Triton pipeline reusing the shared KV latent as both K and V Full analysis: marktechpost.com/2026/08/05/… Technical details: research.meta.ai/blog/introd… Model: developer.meta.com/ai/models… @AIatMeta @Meta_Engineers
4
1,918
Replying to @AIatMeta
“continue with github” was the obvious answer and these mfs chose influencer onboarding.
1
202
Replying to @AIatMeta
2
733
Replying to @AIatMeta
Interesting architecture. Long-horizon agents depend as much on runtime design as model capability. Persistent execution, state management, and reliable recovery become increasingly important as tasks grow more complex.
1
34
Replying to @AIatMeta
Congrats! I'm cheering for you Meta.
81
Replying to @AIatMeta
Yeah, long-running agents also raise the bar for the infrastructure beneath them. Reliable connectivity between models, tools and compute becomes essential when workflows span hours instead of minutes.
71
Replying to @AIatMeta
The orchestration here is cool. It'll be interesting to see how long-term memory layers evolve alongside it.
42
Replying to @AIatMeta
Tools designed for extended engineering efforts reduce the need for constant human guidance. Local event tracking helps maintain clear progress across large repositories.
228
Replying to @AIatMeta
what about benchmarks against @Kimi_Moonshot?
409
Replying to @AIatMeta
Persistent async sub-agents + append-only event log is the real product here — not Spark 1.2. Curious where termination lives: after 1,000+ tool calls / 24h kernel runs, who decides the agent is done vs still "making progress"? In multi-agent loops I've run, liveness fails more often than safety(and time it took
1
52
Replying to @AIatMeta
i've seen persistent agents compound stale assumptions; muse code needs explicit plan invalidation
48
Replying to @AIatMeta
Long-horizon coding agents are becoming genuinely practical. Persistent sub-agents that can plan, implement, and validate across large codebases could make complex software engineering much more efficient. Excited to see Muse Code in action. 🚀
20
Replying to @AIatMeta
Is there no credits for first time users to try out the model?
152
Replying to @AIatMeta
长任务最怕跑偏，验证这步挺关键
131
Replying to @AIatMeta
i also do long-horizon planning but mine is just getting five humans to agree on dinner
276
Replying to @AIatMeta
Meta’s Muse Code: when your coding assistant decides it wants a promotion to full-blown engineer! The edge isn’t another terminal wrapper. It’s persistent background agents that keep context, parallel sub-agents in isolated worktrees, and an event log that survives failures. Built for multi-file features, big refactors, and multi-hour work — not just the next function. Late to the party vs Claude Code, but co-trained model + harness from day one + cheap contributor tier make it worth serious testing. The winners here will be the agents that stay coherent the longest.
2
2
290
Replying to @AIatMeta
yeah now make muse-spark-1.2-contributor availabel in muse code
2
1,510
Replying to @AIatMeta
Hi! While you're building amazing AI tools, many of us are still waiting for our disabled Instagram accounts to be reviewed. My account @jaaavidi has been disabled for over 50 days despite multiple appeals. I'd really appreciate any help getting a human review. 🙏
4
1
206
Replying to @AIatMeta @alexandr_wang
Еще недавно NURM был схожий дизайн на сайте, но теперь там график
2
18
Replying to @AIatMeta
everyone's shipping a terminal coding agent now. the real test isn't the demo, it's whether it survives a messy 200k-line repo
2
44
Replying to @AIatMeta
@AIatMeta A terminal agent with persistent sub-agents changes the unit of work from the file to the outcome. That raises the bar for memory, permissions, rollback, and trust. The interface is a shell. The product is controlled autonomy.
2
327
Replying to @AIatMeta
A coding agent that can keep state while planning and validating multi-file changes is a different tool from a chat window that hands you a snippet and waits for the next prompt. The real test is whether it can finish repo-level work without constant steering.
2
668
Replying to @AIatMeta
please create meta ai app for mac and ide!
2
381
Replying to @AIatMeta
let me use it with Hermes agent via a max subscription please
2
2,090
Load more