MiroThinker is a deep research agent optimized for complex research and prediction tasks. Our latest models, MiroThinker-1.7, achieves 74.0 and 75.3 on the BrowseComp and BrowseComp Zh, respectively.

PyTorch bindings for CUTLASS grouped GEMM.

Fast and memory-efficient exact attention

[CVPR 2026] LongVT: Incentivizing "Thinking with Long Videos" via Native Tool Calling

Sampling profiler for Python programs

verl/HybridFlow: A Flexible and Efficient RL Post-Training Framework

MiroTrain is an efficient and algorithm-first framework research agent.

Train speculative decoding models effortlessly and port them smoothly to SGLang serving.

QQQ is an innovative and hardware-optimized W4A8 quantization solution for LLMs.

a static analytical model for LLM distributed training

Cosmos-RL is a flexible and scalable Reinforcement Learning framework specialized for Physical AI applications.

Efficient Triton Kernels for LLM Training

A PyTorch native platform for training generative AI models

Development repository for the Triton language and compiler

A fast MoE impl for PyTorch

Using megatron style to do TP training.

Open-Sora: Democratizing Efficient Video Production for All

Test the GPU bandwidth of collectives operators like all-reduce, all-gather, broadcast and all-to-all primitives on single-node multi-GPU (2, 4, 8 cards) and multi-node multi-GPU (16 cards) setups,…

Making large AI models cheaper, faster and more accessible

Build a llama fine-tuning script from scratch using PyTorch and transformers API. It needs to support 4 optional features: gradient checkpointing, mixed precision, data parallelism, tensor parallel…

Making large AI models cheaper, faster and more accessible