On October 20, the InternScience team at Shanghai AI Laboratory posted the technical report for Sage-A1.5 on arXiv (2610.28578), with the model card and the full weight family landing on Hugging Face shortly after. It is a 35B Mixture-of-Experts agentic model post-trained from Qwen3.6-35B-A3B, and the team frames it as an upgrade over A1: the model no longer just solves individual tasks, it learns to organize work and revise its own approach mid-execution. Within a day, the Hugging Face paper page had collected 58 upvotes.
Capability composition: three new engines
The report splits the upgrade into three parts. Agentic orchestration: expose a spawn_subagent tool and the model fans tasks out to parallel subagents, each with its own context window and tool access, working without mutual interference before synthesizing findings. Verification mode: enabled through chat_template_kwargs, the model pauses inside its own chain of thought to check which conclusions hold and which search paths are dead, then switches direction in the same turn — no external critic model required. Open scientific discovery: the model moves from answering predefined questions to exploratory research, connecting scientific inquiry with interactive analysis and inspectable artifacts, demonstrated across multidisciplinary scientific applications.
Official benchmarks: 35B weighing in against bigger models
Every number in the model card is vendor-reported; the community has no third-party reproduction yet, so read them as official figures. Sage-A1.5 reaches 95.1 on GAIA, which the team says beats every larger-scale model in its comparison set, including Kimi-K3, GLM-5.3 and GPT-5.6-sol. It posts XBench-DS-2510 at 92.0, IFBench at 81.6, FrontierScience-Olympiad at 81.0, FrontierScience-Research at 48.3 and OmniDocBench at 91.5. Among comparable 35B-class models it ranks first on SWE-bench Pro (61.2), BrowseComp (81.7) and HLE with tools (55.1). Against the previous A1 generation, engineering workloads improved the most: SWE-bench Pro +17.6, SWE-bench Multilingual +14.0, MLE-Lite +21.3, PinchBench +13.1.
Deployment: one GPU and GGUF
The serving story is pragmatic. A single GPU runs a 262K-context vLLM deployment, with multi-token-prediction speculative decoding available (EAGLE under SGLang, mtp config under vLLM). Text-only workloads can skip the vision encoder to free KV cache — the model card tags it as a VLM. Local users get FP8, Q4_K_M, Q8_0 and F16 GGUF editions, all under Apache 2.0.
The interesting part is not any single score but the bet: while frontier models grind forward on scale, this 35B line wagers that orchestration and self-verification buy more than raw size. If the official numbers survive independent reproduction, the entry bar for open scientific agents drops another notch.