Equal token counts don’t mean equal work. Our scheduler uses estimated compute cost to balance work across DP workers and PP microbatches. Within a sequence, later tokens attend to more preceding tokens, making them more expensive to process. Giving each GPU a mix of early and late chunks helps balance that work. This integrates nicely with the rest of our RL stack, including MoE routing replay and the vision encoder.

Another Neon infra tidbit: @periodiclabs accelerated checkpoint conversion ~30x down to just 1 minute, helping us deploy faster to get rapid feedback from our labs. Our models train in Megatron, but SGLang inference consumes weights in HuggingFace format. Traditionally, conversion ran serially: 1. rebuild the entire checkpoint from individual weight tensors 2. convert the checkpoint 3. reshard it to be inference-ready This could take 30 min for a trillion parameter model! Noticing that conversion shouldn't require materializing the full checkpoint, @hsu_byron introduced Fast Resharding: 1. parallelize conversion across Ray actors 2. convert expert chunks directly instead of assembling the full expert tensor in memory We upstreamed Fast Resharding to Miles in PR #1371#. Now even you can deploy trained models to production in a matter of minutes.

Excited to welcome @asadovsky as Harvey’s Chief Research Officer. Before Harvey, Adam co-led post-training at Microsoft AI and Google DeepMind. As a CVP at Microsoft AI, he helped build MAI-Thinking-1, Microsoft’s reasoning model. As part of Gemini’s leadership team he helped train Gemini 1.0 through 2.5, including fine-tuning, RL, data, and evals. His prior work as a Distinguished Engineer at Google spanned Assistant, Search Quality, and Search Infrastructure. I met Adam three years ago when I sent him a cold LinkedIn DM and was surprised he responded. At a time when most dismissed the application layer and legal, Adam was curious and generous with his time. He quickly became someone I regularly turned to for advice on AI as we scaled Harvey over the past three years. When we first met, we were too early to hire someone of his caliber and scale, but I always hoped we’d eventually work together. As Winston and I got to know him better, what stood out even beyond his technical achievements was his character. Despite his incredible technical career, he remains curious, humble, practical, and cares deeply about the teams he builds. We couldn’t think of a better leader to help us build frontier intelligence for the professionals and institutions we serve.

You saw the AI & science. Let's talk about the RL infra it took to build @periodiclabs Neon. To minimize training-inference mismatch in RL, SGLang captures inference's MoE routing decisions for each rollout and we "replay" them while training. In agentic (multiturn tool-use) settings, SGLang exports these router decisions in response to each decoding request, i.e. after each conversation turn. So when *any* data-parallel rank finishes a conversation turn, *all* other ranks must wait until routing data finishes exporting. This slowdown is exacerbated because we export routing decisions from the *entire* conversation rather than just the most recent turn! When @hsu_byron @vwxyzjn discovered this in our Kimi K2 RL setup, they introduced Delta Router Replay: cache previous turns' router decisions on the training client, so you can export only the delta (most recent turn's router decisions) upon each decoding request. Delta Router Replay significantly speeds up our long-context agentic RL runs, and @hsu_byron upstreamed it to SGLang (#24851#) a few months ago.

We check OOD generalization for @periodiclabs Neon and report optimal cost-performance here, too. This indicates useful generalization for our model. But our experimental data also opens up an interesting ML research program. For example, RL tasks to directly predict experimental outcomes given all prior experimental data known up until that date. This is loosely analogous to the often-proposed ML experiment of training AI on all literature before 1905 and then evaluating whether it can derive the theory of relativity. An AI that already knows the answer through pre-training can cheat on RL tasks like this (i.e., it can skip reasoning and simply output a memorized experimental result). A unique, complete, dated system of record makes this work possible and fruitful.

Watch Neon investigate a real sample from our superconductor lab It hypothesizes about crystal structure, reasons about synthesis conditions, and iterates with tools until it finds a physically realistic solution More and more of science will look like this in coming years

Periodic Labs is doing extremely cool work applying frontier AI research to materials science. Here they post-train a model for a X-ray diffraction analysis, beating GPT-6 Astra and boosting performance of Kimi K2.6 by 20x! What's worth noting here is Periodic is doing RL directly with real-world experimental data, which comes with many unique challenges over coding/knowledge-work environments.

Check out our work training custom models on an important part of the materials discovery loop. The scaling direction is pretty clear. We will continue getting bigger and smarter models, these models will have access to better simulations and higher-throughput labs, and they will run increasingly autonomous physical experiments. Periodic will be the best place in the world to do this type of training.