IQuestLab/SAIL

---

Scientific Agentic Intelligence via a Science-Aware Loop

🤗 SAIL Model  |  🤗 SAIL Training Data  |  📄 Paper

Introduction

SAIL is a scientific agent model from IQuest Research, post-trained from Qwen3.6-35B-A3B, with 35B total and 3B active parameters. It searches and analyzes literature, writes scientific code, and uses tools to carry out research tasks that connect evidence, computation, and interpretation.

SAIL's science-aware loop turns gaps observed during scientific task execution into targeted training tasks. Development agents powered by frontier models diagnose these gaps from trajectories, then construct problems and tool-use tasks from scientific papers and code repositories. The capability targets depend on the task: evidence coverage and selection for literature retrieval, domain knowledge for scientific coding, and sustained reasoning across experiments for end-to-end research. Training combines supervised fine-tuning, specialist training, multi-teacher on-policy distillation, and agentic reinforcement learning, followed by reassessment on separate tasks.

We release SAIL and the accompanying training data to support research toward more capable AI scientists.

Performance

SAIL delivers competitive results across scientific workflows alongside substantially larger open-weight models. All models are evaluated in the same benchmark environments, with model-specific inference settings. Higher scores are better; bold values mark the best score in each column.

The parameter comparison uses the unweighted mean of 12 task scores on a 0–100 scale, excluding LitQA2-FullText and counting E2E-Bench Basic and Hard separately.

Literature understanding

Model	Params. (B), total / active	PaperFindings	LitQA-search	ScholarQA-CS2	LitQA2-FullText	ArxivDIGESTables
Ling-3.0-flash	124 / 5.1	18.03	26.67	67.88	91.23	25.45
DeepSeek-V4-Flash-0731	284 / 13	26.46	74.67	75.19	95.08	35.13
Hy3	295 / 21	28.90	73.33	85.63	94.12	32.42
MiMo-V2.5	310 / 15	16.21	34.67	60.85	94.23	27.30
GLM-5.2	744 / 40	40.80	88.00	87.87	90.41	34.21
Ring-2.6-1T	1000 / 63	26.00	50.67	71.17	82.05	25.62
MiMo-V2.5-Pro	1020 / 42	28.05	57.33	75.15	89.06	31.21
LongCat-2.0	1600 / 48	9.69	8.00	41.36	93.75	25.15
BigBang-v1	35 / 3	28.36	56.00	54.32	94.67	28.67
Apodex-1.0-mini	35 / 3	23.77	76.00	74.32	92.00	28.72
Nex-N2-mini	35 / 3	21.78	33.33	46.77	94.67	26.14
Nex-N2.5-mini	35 / 3	12.05	22.67	25.61	91.94	30.94
Agents-A1	35 / 3	22.73	52.00	64.90	95.24	23.28
Qwen3.6-35B-A3B	35 / 3	22.19	48.00	68.62	85.33	25.63
SAIL	35 / 3	33.25	85.33	86.51	91.67	35.24

Code execution, data analysis, and discovery

Model	Params. (B), total / active	DS-1k	SUPER-Expert	CORE-Hard	DiscoveryBench	E2E-Bench	E2E-Bench-Hard
Ling-3.0-flash	124 / 5.1	67.33	31.50	51.35	27.79	63.51	51.99
DeepSeek-V4-Flash-0731	284 / 13	80.56	46.26	72.22	36.75	93.96	86.18
Hy3	295 / 21	80.56	32.87	65.71	35.35	92.51	79.21
MiMo-V2.5	310 / 15	71.89	35.67	54.05	36.50	65.23	50.49
GLM-5.2	744 / 40	76.00	46.66	78.38	37.20	93.77	83.67
Ring-2.6-1T	1000 / 63	52.22	34.06	29.73	26.73	48.31	48.97
MiMo-V2.5-Pro	1020 / 42	68.22	36.66	64.86	44.49	74.83	63.87
LongCat-2.0	1600 / 48	66.67	30.27	51.35	27.88	52.45	42.85
BigBang-v1	35 / 3	67.11	31.54	51.35	33.55	75.00	68.26
Apodex-1.0-mini	35 / 3	68.11	24.83	43.20	31.20	39.94	39.47
Nex-N2-mini	35 / 3	62.70	30.98	62.20	32.07	62.74	53.90
Nex-N2.5-mini	35 / 3	51.10	34.57	64.86	33.21	83.83	70.02
Agents-A1	35 / 3	72.22	30.98	56.80	33.91	32.39	18.28
Qwen3.6-35B-A3B	35 / 3	57.20	28.24	43.20	34.69	58.75	56.06
SAIL	35 / 3	74.30	37.78	67.57	37.48	89.46	77.27

Scientific coding and research

Model	Params. (B), total / active	SciCode	DeepResearch Bench II
Ling-3.0-flash	124 / 5.1	38.19	41.73
DeepSeek-V4-Flash-0731	284 / 13	39.17	43.22
Hy3	295 / 21	38.19	42.54
MiMo-V2.5	310 / 15	27.64	27.46
GLM-5.2	744 / 40	47.57	45.51
Ring-2.6-1T	1000 / 63	41.67	42.84
MiMo-V2.5-Pro	1020 / 42	40.28	41.70
LongCat-2.0	1600 / 48	26.74	35.39
BigBang-v1	35 / 3	41.70	38.55
Apodex-1.0-mini	35 / 3	43.10	37.91
Nex-N2-mini	35 / 3	35.10	41.00
Nex-N2.5-mini	35 / 3	26.83	33.55
Agents-A1	35 / 3	38.19	33.33
Qwen3.6-35B-A3B	35 / 3	39.90	32.27
SAIL	35 / 3	50.35	42.61

Quick Start

Serve SAIL with either vLLM or SGLang in a fresh environment. Both examples enable tool calling and expose http://localhost:8000/v1 under the model name sail.

The commands use eight GPUs and a 262,144-token context limit. Adjust tensor parallelism and context length to fit your hardware. Replace the repository ID with a local model directory to use downloaded weights.

vLLM

pip install -U "vllm>=0.19.0"

vllm serve IQuestLab/SAIL \
  --served-model-name sail \
  --host 127.0.0.1 \
  --port 8000 \
  --tensor-parallel-size 8 \
  --max-model-len 262144 \
  --language-model-only \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

SGLang

pip install -U "sglang[all]>=0.5.10"

python -m sglang.launch_server \
  --model-path IQuestLab/SAIL \
  --served-model-name sail \
  --host 127.0.0.1 \
  --port 8000 \
  --tp-size 8 \
  --mem-fraction-static 0.8 \
  --context-length 262144 \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder

See the vLLM deployment guide and SGLang installation guide for environment setup.

Send a request

curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "sail",
    "messages": [{"role": "user", "content": "Write a Python function to simulate exponential decay and explain how to validate its numerical accuracy."}],
    "max_tokens": 8192
  }'

For tool use, include function definitions in the request's tools field and set tool_choice to "auto". Your application executes the returned tool calls and sends their results back to the model to continue the task. Search services, code execution, and research environments are supplied by the application.

License

SAIL is released under the Apache License 2.0.

Citation

@misc{sailmodelteam2026sailscientificagenticintelligence,
  title         = {Scientific Agentic Intelligence via a Science-Aware Loop},
  author        = {{SAIL Model Team}},
  year          = {2026},
  eprint        = {2610.11451},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2610.11451}
}