Shanghai AI Lab's 397B science agent shipped in July; the paper explaining it landed August 13

Rendering a page and training on the picture is a claim about where scientific content actually lives. Shanghai AI Lab's visual pretraining works from rendered pages rather than text pulled out of PDFs, which keeps figures, tables, equations and layout attached to the prose around them ^[5]. The alternative is familiar to anyone who has pointed an extraction pipeline at a corpus of chemistry papers: table cells arrive in the wrong order and axis labels do not arrive at all. That decision, more than the parameter count, is what changes the class of question you can put to a paper the model has not seen.

The parameter count still deserves a look. The lab's name says 397B, while the Hugging Face listing says roughly 404 billion parameters and about 810 GB of files ^[11]. Four hundred and four billion values at two bytes each is 808 GB, so what is on offer is unquantized 16-bit weights rather than a packed release ^[19]. Label and listing differ by about 7 billion parameters, or 1.8 percent ^[22]. More usefully: 810 GB of weights spread over 80 GB accelerators occupies eleven of them before a byte of KV cache is allocated ^[20].

What the training pipeline is built for is harder than what the published scores measure. It combines supervised fine-tuning, reinforcement learning across scientific domains, reinforcement learning inside interactive agent environments and on-policy distillation, with a separate branch that forecasts time-series values numerically instead of spelling them out as text tokens ^[8][9]. The evaluation, meanwhile, capped text reasoning at 256,000 tokens and multimodal tasks at 64,000, and the authors say those limits describe the harness rather than evidence the model can finish an open-ended research programme without losing direction ^[7]. Serving that 256,000-token window sits on top of the eleven-device floor, not beside it. The breadth being attempted here follows chief scientist Bowen Zhou's stated thesis that a general model should keep broad capability while acquiring deeper expertise in individual fields ^[18].

The paper trailed the weights by about four weeks ^[21]. That ordering is now ordinary for open-weights releases, and the cost falls on whoever moved first: the earliest deployers ran a system whose training recipe and evaluation caveats were not yet written down ^[4]. Most teams will meet the model through the lab's OpenAI-compatible API in any case ^[12], which is the same arrangement they already have with closed models. Downloadable weights only change the relationship for the people who can hold them.