Microsoft AI details Rocket, an in-house distributed reinforcement learning framework using SGLang to train its MAI-Thinking-1 model

and if you're done with this thread and still want to read more about this report, pease take a look at the goat @stochasticchasm recap

this was an insanely good read, i think this is the most detailed report i've read at this scale in some aspects. i really hope MAI continues releasing those tech reports, thanks a lot to the team for this gift 🥹 https://microsoft.ai/wp-content/uploads/2026/06/main_20260602_2.pdf#page=81.11

microsoft uses SGlang wow

will conclude by this, 40% higher throughput per Watt (or is it different from "rack power budget"?) is pretty impressive and bullish on microsoft chips

good infra numbers about the final training run, love the transparency here

@stochasticchasm and the one from @nrehiew_ 🐐

My takeaway from the RL part (something perhaps everyone agrees) is that the initial iterations of training a reasoning model from scratch are overwhelmingly a stability‑engineering problem. Now they have the infrastructure in place, the key was also getting the data/tasks modelling (many times I see folks directly exploring how to improve the algorithms, thinking through the structure of the tasks). Now, it should give them a lot of freedom to explore in the algorithmic space.

@eliebakouch Very happy that you liked it :D

@eliebakouch @stochasticchasm @threadreaderapp please unroll!

@eliebakouch @stochasticchasm Another interesting fact is that they start their RL from a checkpoint that hasn't been exposed to reasoning trajectories!

@slime_framework router rebalance lag idled 20% of gpus when rollout waited on inference

@slime_framework Glad to see Rocket's async design gaining traction. The real challenge is integrating this at scale within existing workflows.

@eliebakouch @TheZachMueller many such cases

@eliebakouch @TheZachMueller What else?

@northead @eliebakouch @stochasticchasm @northead Hi, please find the unroll here: https://threadreaderapp.com/thread/2061965825037254947.html Talk to you soon. 🤖

@harold_matmul @eliebakouch Awesome work, enjoyed going through the report! Had qq about agentic evals: For terminalbench, seems you didn't use Harbor/Terminus-agent rather used a ReAct-loop with tool dispatches. This might make numbers less comparable, was there a reason to not use Terminus-agent directly?