/Tech30d ago

Nan Jiang and Modal's Charles Frye call for collaboration on scaling open-source reinforcement learning infrastructure

They target delta compression and weight synchronization challenges

203322111656.8K

#19

Original post

Ben (no treats)#1623

Nan Jiang@nanjiangwill

At @modal, we're working to make sure OSS RL frameworks have all the techniques necessary to train frontier open-weights models. Delta compression is key, but the job's not done. There are still lots of open problems around weight sync, auto-scaling, & cross-cluster training.

My DMs are open!

slime@slime_framework

@FireworksAI_HQ + @cursor_ai highlighted why delta-compressed weight sync matters for RL at frontier scale.

slime brings this capability to OSS: lossless delta sync for Megatron ↔ SGLang disaggregation — ship deltas, not full checkpoints.

This is another step toward a fully open-source stack where rollout/inference and training are truly decoupled and deployed separately.

PR: https://github.com/THUDM/slime/pull/1806

12:56 PM · May 30, 2026 · 37.4K Views

Sentiment

Users praise Modal's delta-compressed weight sync for open-source frontier RL training because it builds on an amazing battle-tested framework from the slime_framework community.

Pos

100.0%

Neg

0.0%

2 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Related links

1806

GITHUB.COMVia

Posts from X

Most Activity

VIEWS14.3KBOOKMARKS15LIKES80REPLIES8RETWEETS3

Peyton Walters@peywalt

true story: nan was not allowed to leave his @modal nyc onboarding until he got delta compression working in slime.

nan had delta compression working within 1.5 weeks of joining modal and was allowed to return to sf.

Nan Jiang@nanjiangwill

My DMs are open!

29d14.3K8015

Charles 🎉 Frye @ AIEng World's Fair@charles_irl

why use many bytes when few do trick?

Nan Jiang@nanjiangwill

My DMs are open!

30d4.6K364

Ben (no treats)@andersonbcdefg

@charles_irl when me jensen they see

Charles 🎉 Frye @ AIEng World's Fair@charles_irl

why use many bytes when few do trick?

30d51440

Nan Jiang@nanjiangwill

Huge thanks to the @slime_framework community for making an amazing, battle-tested RL framework!

I think we are well-positioned at Modal to help users deploy slime. On our infrastructure, train/inference disaggregation can pair naturally with elastic scaling, so rollout capacity is neither wasted nor bottlenecked.

30d772

Charles 🎉 Frye@charles_irl

@peywalt @modal this was nan-negotiable

29d392

Ashton Chew@iamashtonchew

@nanjiangwill @modal When the Greek God of VLM speaks, I listen

30d33