NeMo Megatron Bridge

Documentation | Supported Models | Examples | Contributing

📣 News

• [11/05/2026] ERNIE-5.5-Flash-Base is now merged on main! See the examples README for sft training details.

• [10/16/2026] Reflection AI Beam-1.0-501B-A23B is now supported! Megatron Bridge provides HF↔Megatron checkpoint conversion, inference, and BF16 H200 recipes for pretraining, SFT, long-context SFT, and LoRA PEFT. See the examples README and verification card for verified configurations and current limitations.

• [10/15/2026] NVIDIA Nemotron 3.5 Super VL is released! Megatron Bridge supports the 120B-A12B vision-language model with HF↔Megatron checkpoint conversion, image inference, and pretraining, SFT, and LoRA recipes, plus text-only conversion and training workflows. See the model documentation and verification card for supported workflows, verified configurations, and current limitations.

• [10/15/2026] Muse-Spark-Open-816B-A42B is now supported! Megatron Bridge adds checkpoint conversion, inference, and GB200 training recipes for the 816B-total / 42B-active vision-and-audio-language MoE model: pretraining, full SFT, long-context SFT at 32K and 64K with context parallelism, LoRA, and image-text SFT and LoRA, with HybridEP MoE dispatch. See the examples README and the verification card for details.

• [08/19/2026] Megatron Bridge 0.6.0 released! Highlights include GLM-5.2 with cuDNN fused DSA for 128K THD-packed context-parallel training, new LLM/VLM/diffusion models, more reliable VLM data and sequence packing, improved checkpoint and LoRA workflows, unified training and conversion launchers, and refreshed performance recipes. Huge thanks to our community contributors! See the full release notes.

• [08/11/2026] NVIDIA Nemotron 3.5 Lightning is released! Read the technical blog here. See the Nemotron 3.5 Lightning README for details.

• [07/31/2026] K-EXAONE-2 is now supported! Day-0 support for LG AI Research's K-EXAONE-2 (750B-A37B MoE) is now available on the k-exaone-2 branch. Users can convert checkpoints, run inference, fine-tune with SFT or PEFT (LoRA), and use the NVFP4 inference-optimized checkpoint released alongside the model. See the branch documentation for setup instructions and examples.

• [07/17/2026] LongStraw (MinT-2M) is a long-context RL research system from MindLab Research that uses Megatron Bridge's modeling and LoRA capabilities. LongStraw explores resident-prefix, response-only GRPO by capturing a shared prompt once and replaying only the trainable response branches, and reports a 2.1M-token GLM-5.2 execution path across 32 NVIDIA H20 GPUs. The broader MinT system behind this work was also used to post-train Macaron-V1-Preview, MindLab's 749B Mixture-of-LoRA agent model derived from GLM-5.1. Learn more in the paper.

This is the beginning of the README. Read the rest at the original. Read the full README ↗