PyTorch post adds TorchTitan tuning for Muse Spark Open

On 15 October 2026 at 17:55 UTC, the PyTorch blog published Muse Spark Is Now an Open-Weights Model. The page records a later edit the same day at 19:18 UTC. The weights release itself is Meta releases Muse Spark Open weights under a gated license. This note is the PyTorch-stack follow-up.

The blog says the weights are on Hugging Face under the Muse Model License. That is open weights under Meta’s license. The earlier post already reported Meta’s claim of Helion kernel speedups in SGLang of 1.03 to 3.95 times on NVIDIA and AMD GPUs. The PyTorch post names the hardware, and it adds Helion megakernels and fine-tuning in TorchTitan.

What the kernel numbers refer to

The blog says Muse Spark Open ships with Helion implementations of operations such as RoPE, RMSNorm, and quantization. It claims speedups of 1.03 to 3.95 times against custom kernels written in hardware-specific languages, with NVIDIA CUDA and the CuTe DSL as the examples, on an NVIDIA GB300. It then says those kernels were portable to AMD and scored similar performance improvements on an AMD MI350, which the same paragraph also writes as “Mi350,” with an overall 3% improvement in end-to-end throughput.

A later paragraph widens the kernel range. It says Muse Spark Open ships with Helion GPU kernels in SGLang, delivering 1.03 to 3.95 times kernel-level speedups on NVIDIA and AMD GPUs and a 3% end-to-end speedup on AMD MI350. The 1.03 to 3.95 times figure in the first paragraph is a GB300 comparison with CUDA and CuTe kernels. The 3% figure is end-to-end, on MI350. Both are the blog’s claims. The earlier weights post did not split them this way.

The same section says Helion megakernels for Muse Spark Open shipped in SGLang. Written in Helion, they fuse several operations into one persistent kernel for low-latency inference, and they can be autotuned for shapes and hardware. The blog also says Helion joined the PyTorch Foundation in April 2026.

TorchTitan

The post says that starting on 15 October, TorchTitan adds a model definition for Muse Spark Open. It says TorchTitan has native multi-turn supervised fine-tuning and LoRA, and reinforcement learning through TitanRL, so a user can fine-tune this checkpoint. That support is the blog’s statement. The TorchTitan repository was not opened to confirm the files landed.

A reader of the weights post can treat the new facts as these: a GB300 kernel range versus CUDA and CuTe, a 3% end-to-end claim on MI350, SGLang megakernels, and supervised fine-tuning, LoRA, and TitanRL inside TorchTitan. The license is still the Muse Model License.