# TorchTitan adds new model support and reinforcement‑learning features

> PyTorch’s TorchTitan platform now trains the latest open‑weight models and includes a unified RL stack called TitanRL.

Oossa · 2026-10-21 · https://oossa.com/en/torchtitan-adds-new-model-support-and-reinforcement-learning-features

TorchTitan, the PyTorch‑native training platform, announced several upgrades. It now supports the latest open‑weight models such as Muse Spark Open, Qwen3‑VL, Kimi K2.7, Kimi K3 and DeepSeek‑V4, all integrated with its 5‑dimensional parallelism. The same release adds TitanRL, a reinforcement‑learning framework that reuses the exact pretrained model for both pretraining and inference, removing the need for separate codebases. Performance improvements include faster kernels, better memory handling and new communication features that speed up large‑scale training.

## The facts

- New model list includes Muse Spark Open, Qwen3‑VL, Kimi K2.7, Kimi K3, DeepSeek‑V4
- TitanRL enables identical model code for pretraining, post‑training and inference

## Why it matters

Researchers can adopt the newest models and run RL experiments without rewriting code, saving time and reducing bugs.

## Sources & references

1. [TorchTitan Update: New SOTA Models, Performance at Scale, TitanRL](https://pytorch.org/blog/torchtitan-update-new-sota-models-performance-at-scale-titanrl/) – PyTorch

Last updated: 2026-10-21