THUDM has released Slime as a fully open-source framework that handles the post-training stage of large language models through reinforcement learning scaling and online preference optimization, with the same stack used to finish GLM-5.2 training in roughly two days.
Thanks for the support!
A small note: slime has supported not only OPD, but the full RL + OPD post-training workflow since GLM-4.5.
More to come for scalable agentic RL infra.
Incredible how Z. ai literally has their RL infrastructure open source.
The entire OPD post-training of GLM-5.2 took on this slime platform took ~2 days.
https://github.com/THUDM/slime
Native Megatron-LM and SGLang hooks plus support for synchronous or asynchronous modes let developers plug in their own verifiers, environments, and data workflows without rewriting core training plumbing.
The repo shows use across GLM variants plus Qwen, DeepSeek V3, and Llama 3, yet no adoption counts, contributor activity, or scaling benchmarks beyond the GLM-5.2 timeline have been shared.
Many users are excited about THUDM open-sourcing the Slime RL platform for GLM post-training because of its efficiency, versatility, flexibility with custom backends, and lowered reproduction barriers for researchers and small teams.
No Digg Deeper questions have been answered for this story yet.
@didier_lopes > The entire OPD post-training of GLM-5.2 took on this slime platform took ~2 days.
what
Incredible how Z. ai literally has their RL infrastructure open source.
The entire OPD post-training of GLM-5.2 took on this slime platform took ~2 days.
https://github.com/THUDM/slime
@Infopulsed This is probably why everyone is so paranoid about distillation - it seems it's incredibly efficient.
This is from Qwen3's technical report.
@alexbastian_ai They do shed some light into the type of data they use for pre-training
this is actually really cool, fully open source. RL framework behind GLM-5.2, GLM-5.1, GLM-5, GLM-4.7, GLM-4.6, and GLM-4.5, validating the full post-training loop.
@teortaxesTex @didier_lopes this is after training all of the expert variants and collapse them back into the same base model? seems plausible to me the signal is much more dense in that phase
@didier_lopes Nice, now open source the pre-training code and all the datasets.
@vikhyatk @didier_lopes yeah the word "entire" was throwing me off but for OPD I guess makes sense, if they encounter no problems
@teortaxesTex @didier_lopes 200 experiments to figure out the optimal policy. and then 2 days to do the final training run
@didier_lopes yup, it contains their exact Megatron +sglang setup, you can even checkout the some of the older GLM versions in other branches.
@didier_lopes they are on X @slime_framework
@Pran_Ker amazing.
Not even 1.0.0 yet, we are so early
@didier_lopes GLM-5.2’s 2-day fine-tune on an open RL platform is a solid signal. most RL frameworks take weeks even with larger teams. this could reset expectations for solo or small-team builders.
@didier_lopes Thanks for supporting @slime_framework ❤️
https://github.com/THUDM/slime
@didier_lopes Yo this is crazy!! What???
@didier_lopes two days for post training is wild efficiency, feels like the kind of flex nobody expected from them yet
@didier_lopes 开源到这份上，复现门槛真降了
@didier_lopes @cobaltdigital33 just found this!
@vikhyatk @teortaxesTex I figure it's like when DeepSeek-V3 came out with the $5.5 M run that excluded everything else.
The only datapoint I could find on this is from Qwen3-235B-A22B technical report https://arxiv.org/pdf/2505.09388
so the 2 days number doesn't sound as crazy as it did
