Muse Code and Muse Spark 1.2

We're excited to merchandise Muse Code (beta), a terminal coding supplier powered by Muse Spark 1.2, our newest model. This marks our adjacent measurement toward the frontier, pinch larger and overmuch much tin models connected the way.

Install Muse Code connected macOS aliases Linux:

Muse Code takes connected analyzable package engineering tasks crossed ample repositories: readying changes, penning code, and validating the results. It tin coordinate aggregate persistent subagents for each task, solving difficult problems faster, much accurately, and pinch little intervention.

Muse Code

Async Background Agents

Muse Code operates pinch a elemental supplier loop positive a group of async inheritance agents to heighten the main agent's capability. These specialized inheritance agents stay progressive passim each session, alternatively than being spawned for individual tasks, helping debar redundant accusation gathering. They transportation retired adjacent steps and take erstwhile to pass backmost to the main agent. Their persistence reduces latency and the request for steering connected difficult, multi-step tasks.

Runtime Design

Muse Code uses a section arena log successful which each exemplary call, instrumentality run, approval, and edit is appended. This azygous root of truth makes the runtime replay-exact and restart-safe: aft a crash, the supplier tin resume precisely wherever it stopped. That expertise lets Muse Code return connected long-running tasks without being derailed by failures.

Bundled Skills

Muse Code ships pinch respective default skills. /plan turns a task into an approval-gated plan, /grill stress-tests that scheme until it holds up, and /goal useful toward successful completion of the specified objective.

Muse Spark 1.2

Muse Spark 1.2 is simply a coding-focused update to Muse Spark 1.1, pinch improvements successful codification generation, analyzable debugging, codebase understanding, and end-to-end developer workflows. In Muse Spark 1.2, we importantly scaled up training compute connected coding tasks while expanding training situation diversity. The exemplary besides maintains its spot successful different cardinal areas for illustration wide agents.

For much specifications astir our evaluations, spot our report.

Co-Training With Muse Code

We co-trained Muse Spark 1.2 pinch Muse Code to guarantee the exemplary exhibits its champion capacity and coding usability erstwhile paired together. The training included rejection sampled harness trajectories and look optimizations for goals, compaction, and subagents, alongside the integration of the Muse Code toolset to maximize harness compatibility.

Long-Horizon

Muse Spark 1.2 was extensively trained connected long-horizon coding tasks, including whole-repository generation, ample end-to-end projects, and auto-research. It leverages readying to series work, extremity conditioning to support direction, and discourse compaction to clasp the knowledge needed to prolong progress.

Self-Improvement

We besides utilized Muse Spark 1.1 to make challenging coding environments and instruction-following templates. The exemplary past graded campaigner solutions connected really good they satisfied those requirements, producing a scalable training dataset for Muse Spark 1.2. This self-improvement loop helped Muse Spark 1.2 travel analyzable instructions much precisely than its predecessor.

Case Study: Kernel Optimization

We tested the model's expertise to iteratively optimize GPU kernels complete 1,000+ instrumentality calls (up to 24 hours). Leveraging Muse Code's agentic coding environment, the exemplary writes, compiles, profiles, and progressively improves kernel capacity comparative to a provided baseline implementation. We benchmarked connected KDA and MLA kernels for NVIDIA Hopper GPUs. The supplier continues to execute important improvements complete the provided baseline implementation.

The baseline is the FLA Triton implementation of KDA. Models were prohibited from importing third-party kernel libraries specified arsenic FLA directly; instead, they had to use specialized kernel-optimization knowledge to instrumentality the algorithm successful Triton, alternatively than wrap existing implementations. Muse Spark 1.2 paired a chunk-parallel mentation kernel pinch a sequential inter-chunk scan, combining modular fusion and tiling pinch KDA-specific optimizations specified arsenic re-centering the gated cumulative decay astatine the chunk midpoint.

We benchmark against a PyTorch reference implementation astatine batch size 1, number of heads 64, series magnitude 8192, and latent magnitude 512. Muse Spark 1.2 designed a two-kernel Triton pipeline for this workload, combining kernel fusion and tiling pinch MLA-specific optimizations specified arsenic reusing the shared KV latent arsenic some K and V.

Availability

Muse Spark 1.2 is disposable coming successful Muse Code and successful Meta Model API pinch expanded world access. We person a batch connected the horizon, including caller harness features and much powerful models. We can’t hold to spot what you build!

Get started pinch Muse Code