IQuest posted SAIL. The SciCode 50.35 lead is the authors’

What was posted on 8 and 9 October?

The arXiv Atom record we opened lists 2610.11451v1 as published at 08:04:22 UTC on 8 October 2026, primary class cs.CL, with cs.DL and cs.IR secondaries. The comment field says “16 pages, technical report.” Authors start with SAIL Model Team; the HTML marks Bryan Dai as corresponding author and Chi Liu as technical lead.

The public weight repo did not exist on the paper’s posting day. Hugging Face’s model API for IQuestLab/SAIL, retrieved 9 October, gives createdAt 2026-10-09T06:21:20Z, lastModified 2026-10-09T06:51:04Z, sha c3c8d2fbaa928e3f43fe3dfeae998eb07da1948f, gated false, license apache-2.0, and tags that include sail, science, mixture-of-experts, and arxiv:2610.11451. The dataset API for IQuestLab/SAIL-Training-Data gives createdAt 2026-10-09T06:21:40Z and lastModified 2026-10-09T06:51:44Z.

This is a scientific-agent post-train, not a new sequencer or a new literature index. For another Ai2 scientific-writing drop on this site, see AstaBrief 8B. AstaBench, which SAIL uses for most of its tables, is the same Allen Institute suite that paper cites.

What is the science-aware loop?

The abstract says SAIL is developed through a science-aware improvement loop: agents built on frontier AI models analyze the student model’s task failures and construct training tasks that address the gaps. Diagnosis looks at search and evidence selection on literature tasks, scientific assumptions and reasoning on coding tasks, and planning and revision on longer investigations. The agents draw on paper collections and scientific code repositories to build problems, interaction trajectories, and executable tasks with the required environments and tools.

Section 2 repeats that loop over multiple development cycles. Literature failures are split into exploration versus selection. Coding failures are traced to concepts, assumptions, and whether the model notices a bad computational output — the paper’s Figure 3 is a forward-Euler orbit that runs without errors while energy drifts. End-to-end failures are lost objectives, repeated dead ends, or unused intermediate findings. The authors say they validate generated material against sources and task-specific checks, including execution where applicable.

The technical-contributions list is three items: the open 35B / 3B model, the loop, and the training-plus-infrastructure recipe. “We release the model and most of its training data” is the paper’s grant sentence. The dataset card is more specific about what “most” excludes.

How do they train the student?

Section 3 starts from Qwen3.6-35B-A3B. Supervised fine-tuning mixes general instruction data with the scientific tasks from the loop and produces the shared checkpoint. Specialists then train on focused distributions with extra SFT, RL, or both. Multi-teacher on-policy distillation, citing Ma et al. 2026, has the student generate its own trajectories while a routed specialist scores every student token under the same context, including tool observations. Agentic RL then uses task verifiers on executable families. For the MoE shape this starts from, see our mixture-of-experts explainer.

Section 4 says the on-policy stack is built on verl: a training engine, a rollout engine, Uni-Agent adapters into sandboxes, and a gateway that records token-in/token-out IDs. Long episodes that exceed one context window are split when the harness compacts history, but they keep one episode identity and one terminal reward. RL and MOPD share that collection path.

The Hub README’s serve recipes match a language-model-only deployment: vLLM 0.19 or newer, or SGLang 0.5.10 or newer, eight GPUs, a 262,144-token context, the Qwen3 reasoning parser, and the qwen3_coder tool-call parser. The vLLM command passes --language-model-only. Search services, code execution, and research environments, the card says, are supplied by the application.

Which scores are the authors’?

Section 5 says every compared model was tested in the same benchmark environments with that model’s recommended inference settings. AstaBench covers literature, code execution, data analysis, and end-to-end research; SciCode and DeepResearch Bench II are reported beside it. The comparison set is the Qwen3.6-35B-A3B base, five other 35B models, and eight larger models from 124B to 1.6T total parameters.

On SciCode the authors print 50.35 for SAIL, 10.45 points above the base model’s 39.90 and 2.78 points above 744B GLM-5.2 at 47.57. On ArxivDIGESTables they print 35.24, just ahead of DeepSeek-V4-Flash-0731 at 35.13. The unweighted mean of twelve task scores, excluding LitQA2-FullText and counting E2E-Bench Basic and Hard separately, is 59.76 for SAIL, 59.15 for DeepSeek-V4-Flash-0731, and 63.30 for GLM-5.2. SAIL is second on that aggregate in their figure.

Against the base model they report gains on all 13 evaluations, largest on LitQA-search (+37.33), E2E-Bench Basic (+30.71), CORE-Hard (+24.37), and E2E-Bench Hard (+21.21). Against the other 35B models they say they lead 12 settings; LitQA2-FullText is the exception, where Agents-A1 is 95.24 and SAIL is 91.67. Section 6.2 names evidence interpretation, repository execution, and longer investigations as remaining gaps. We did not rerun any row.

What can you download, and under what terms?

The model README we opened points to the paper, the training-data repo, Apache License 2.0, and the vLLM and SGLang commands above. Fifteen model-*.safetensors shards plus tokenizer and processor files are listed on the model API. The card’s base_model field is Qwen/Qwen3.6-35B-A3B. That is an Apache-2.0 post-train of an existing MoE, not a from-scratch public pretrain. For that distinction, see open weights versus open source.

The dataset README states 171,586 scientific training examples for supervised fine-tuning and prints the five-way split: 157,742 scientific-coding, 4,966 research-reasoning, 3,702 literature-analysis, 3,477 research-synthesis, and 1,699 end-to-end-discovery. Records are Parquet under sail-code, sail-search, and sail-e2e. Each row has an id, a messages list, optional tool schemas, and metadata. The same card says this repository provides text conversations and tool definitions; executable tools and task environments are not included.

usedStorage on the dataset API is 1,260,454,735 bytes. The model API’s usedStorage is 70,236,189,701 bytes. Those are Hub accounting figures, not a load we performed. The paper’s “most of its training data” sentence and the dataset’s missing environments are different statements. Do not collapse them.

What did we not run?

This is an evidence review of the 8 October abstract, Atom record, and HTML, and of the 9 October Hugging Face model and dataset API records and README files. We did not download the 70 GB shards, load the 171,586 Parquet rows, stand up vLLM, or rerun SciCode, AstaBench, or DeepResearch Bench II. Treat the loop as what the PDF describes, the percentages as what the authors printed, and the Hub drop as an Apache-2.0 weight-plus-SFT-text release without the execution environments.