Agents-A1
modelYour notes
"Scaling the Horizon, Not the Parameters: Reaching Trillion-Parameter Performance with a 35B Agent" — InternScience's agentic VLM: a 35B-A3B MoE (hybrid linear/full attention, image-text-to-text, Apache-2.0) trained on trajectories averaging ~45K tokens via a three-stage recipe: full-domain SFT, domain-level teacher training, then multi-teacher multi-domain on-policy distillation with heterogeneity-aware optimization. Card-reported results claim SOTA or best-in-class against GPT-5.5, DeepSeek-V4-Pro, and Kimi-K2.6 on agentic and science benchmarks — GAIA 96.0, BrowseComp 75.5, HLE-with-tools 47.6, IFBench 80.6, SciCode 44.3. Strong first-week traction (194 HF likes).
It is a post-train, not a from-scratch model. The card does not name the base, but the paper does ("We fine-tune from Qwen3.5-35B-A3B"), and the config.json is Alibaba's (Qwen3_5MoeForConditionalGeneration, 40 layers, hidden size 2,048, 256 experts with 8 active, 248,320 vocabulary). The 35B / 3B-active count therefore belongs to Qwen3.5 and is not recorded as Shanghai AI Lab's.
Sage-A1.5 (weights 15 October 2026, technical report arXiv 2610.28578 on 20 October, Apache-2.0, with FP8 and GGUF builds) is the second generation of the line, and the card names its lineage outright: "a 35B Mixture-of-Experts agentic model based on Qwen3.6 [Qwen3.6-35B-A3B] and Agents-A1". Its config is the same Qwen3.5/3.6-MoE shape. The report adds three capabilities on top of A1: agentic orchestration (dispatching subagents with isolated contexts in parallel), an opt-in verification mode in which the model reviews its own progress mid-trajectory with no external critic, and open scientific discovery case studies. Training is multi-domain SFT across Claude Code, Codex and internal harnesses ("several hundred million supervised tokens"), then outcome–process post-training with on-policy teacher feedback. One teacher, a Qwen3.6-35B-A3B-based MLE specialist, lifted MLE-Bench Lite medal rate from 37.9% to 72.7%, the same figures reported for the separately released InternScience/MLExpert checkpoint. The card claims Sage-A1.5 posts the best scores among the compared models, frontier systems such as Kimi-K3, GLM-5.3 and GPT-5.6-sol included, on XBench-DS-2510 (92.0), IFBench (81.6), PinchBench (90.0) and FrontierScience-Olympiad (81.0). Among ~35B models it claims BrowseComp 81.7 (82.8 with orchestration), SWE-bench Pro 61.2, AA-LCR 81.0 and HLE with tools 55.1. GAIA came in at 95.1, slightly below A1's 96.0. Against A1 the gains are on engineering: SWE-bench Pro +17.6, SWE-bench Multilingual +14.0, MLE-Lite +21.3. All figures are self-reported; Artificial Analysis has scored neither checkpoint.
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| GAIA | 96.0 | — |
| BrowseComp | 75.5 | — |
| HLE (with tools) | 47.6 | — |
| IFBench | 80.6 | — |
| SciCode | 44.3 | — |
Variants
| Name | Parameters | Notes |
|---|---|---|
| Agents-A1 | — | 2026-06-26; 35B total / 3B active MoE VLM, fine-tuned from Qwen3.5-35B-A3B |
| Sage-A1.5 | — | 2026-10-15 (report arXiv 2610.28578, 2026-10-20); based on Qwen3.6-35B-A3B and Agents-A1; BF16, FP8, GGUF; card: XBench-DS-2510 92.0, IFBench 81.6, GAIA 95.1, BrowseComp 81.7, SWE-bench Pro 61.2, HLE w/ tools 55.1, AA-LCR 81.0 |