NeMo RL v0.8.0 Roadmap
- Dominant language
- Python
- Stars
- 2k
- Forks
- 561
- Avg merge
- 4d 5h
- Merged PRs (30d)
- 145
Description
# NeMo RL v0.8.0 Roadmap
**ETA:** 9/30/2026
This is a community-facing snapshot of the work targeted for the NeMo RL v0.8.0 release. Scope and status may evolve as implementation progresses.
---
## 1. Training - Model support
| Feature | Status | Link |
| -- | -- | -- |
| GLM-5.2 support for AutoModel and Megatron Core | WIP | NVIDIA-NeMo/RL#3172 |
| DeepSeek V4 support for AutoModel | WIP | NVIDIA-NeMo/RL#2331 |
| Nemotron Nano v3.5 support for AutoModel | WIP | PR NVIDIA-NeMo/RL#3383 |
| MiniMax M3 support for AutoModel | WIP | NVIDIA-NeMo/RL#3329 |
| Ultra support | Planned | — |
---
## 2. Low-precision training and quantization
| Feature | Status | Link |
| -- | -- | -- |
| MXFP8 convergence and benchmark milestones | WIP | NVIDIA-NeMo/RL#3431 |
| MXFP4 / NVFP4 recipe support | WIP | NVIDIA-NeMo/RL#3432 |
| Per-token FP8 and NVFP4 quantization | Planned | — |
---
## 3. Inference - SGLang backend
| Feature | Status | Link |
| -- | -- | -- |
| SGLang support with Megatron Core | WIP | PR NVIDIA-NeMo/RL#3190 |
| Non-colocated SGLang with Megatron refit, MXFP8, and fault tolerance | WIP | PRs NVIDIA-NeMo/RL#3190, [#3188](), [#3187](), [#3189]() |
---
## 4. Algorithms and data
| Feature | Status | Link |
| -- | -- | -- |
| PPO with non-colocated generation | WIP | PR NVIDIA-NeMo/RL#3262 |
| Context compaction | Planned | — |
| Single-controller asynchronous GRPO | WIP | NVIDIA-NeMo/RL#2625 |
| Diffusion RL | WIP | NVIDIA-NeMo/RL#3233 |
| Cross-tokenizer distillation | WIP | NVIDIA-NeMo/RL#1827, draft PR NVIDIA-NeMo/RL#3237 |
---
## 5. Platform support
| Feature | Status | Link |
| -- | -- | -- |
| NeMo RL container support and validation for GB300 | WIP | NVIDIA-NeMo/RL#3492 |
Contributor guide
Research direction
Review the roadmap sections and the linked issues and pull requests for model support, quantization, SGLang, algorithms, and platform support. No file, test, entry point, or acceptance criteria is identified, so completion cannot be determined from this issue alone.
Written by the indexing model from the issue text.
Assessment
- Tech stack
- python
- Domain
- ai, machine-learning
- Issue type
- Documentation
- Difficulty
- 5/5
- Estimated time
- Over a week
- Activity status
- Stale
- Clarity
- Needs clarification
- Newbie friendliness
- 15/100