Sage-A1.5: Open 35B agent model uses subagents and self-checking for research

What this is about

Shanghai AI Laboratory (team InternScience) has released the model Sage-A1.5 on Hugging Face. It is an "agentic" language model: it's meant to do more than answer questions. It plans tasks across many steps, uses tools such as search or code execution, and adjusts its approach along the way. The weights are available for download under the Apache 2.0 license, also on ModelScope. There is also a technical report on arXiv, code on GitHub and a homepage.

Technically, Sage-A1.5 is a Mixture-of-Experts (MoE) model. It consists of many specialized sub-networks ("experts"), and only a small share of them is active for each token. That saves compute. It is based on Qwen3.6-35B-A3B and the team's own predecessor, Agents-A1. The model card says "35B"; Hugging Face counts 36B parameters (roughly speaking, the tunable values learned during training). The model is also tagged as a vision-language model (VLM), so it can process images too.

The three core capabilities

The developers name three focus areas:

• Agentic orchestration: The model breaks a task into independent sub-questions and sends out subagents that work in parallel. Each subagent has its own context window (its own "working memory" for the conversation history) and its own tool access. That keeps the subtasks from interfering with each other. The main model then combines the results.
• Verification: An optional mode in which the model pauses within its own chain of thought, checks what is established and what has failed, and changes direction in the same turn. There is no external checker.
• Open scientific discovery: The model is meant to explore and research rather than just answer set questions, and to deliver traceable intermediate results.

Other listed features include native function calling, scientific and domain reasoning, and precise following of instructions with multiple constraints.

Benchmarks (developer-reported)

The comparison covers similarly sized models (Agents-A1, Qwen3.6-35B-A3B, Ornith-1.5) and larger models (Kimi-K3, GLM-5.3, GPT-5.6-sol). The figures come from the developers. For competitors, they took values from those models' technical reports and measured missing ones themselves.

The developers say they achieve the best overall scores on XBench-DS-2510 (92.0), IFBench, PinchBench (90.0), FrontierScience-Olympiad (81.0), FrontierScience-Research, CMT (80.0) and OmniDocBench (91.5). On GAIA, the model scores higher than every larger comparison model, though its predecessor Agents-A1 reaches 96.0 there. Compared with Agents-A1, Sage-A1.5 improves sharply on engineering tasks: SWE-bench Pro +17.6, SWE-bench Multilingual +14.0, PinchBench +13.1, MLE-Lite +21.3.

On many tests, though, it clearly trails the big models. Examples are BrowseComp without orchestration (81.7 versus 91.2 for Kimi-K3), MLE-Lite (65.2 versus 77.3), ChemCOTBench (66.5 versus 82.5) and CritPt (20.0 versus 32.3). On MatTools (42.6) and HiPhO (45.3), even Agents-A1 comes out ahead. For the SWE-bench series, the developers used the mini-SWE-agent scaffold, temperature 1.0, 200K context and a fully offline environment.

How to use the two extra modes

Both modes are off by default.

Orchestration: You register a tool called spawn_subagent. The model calls it when a task can be split into independent investigations. Each subtask needs task, context and expected_output, because subagents can't see the main conversation. They return JSON with an answer, status, evidence and open issues. If they can't find anything, they reply "UNKNOWN" instead of guessing. By default, 5 tasks per call and 20 turns per subagent are allowed. Your own harness, not the model, runs the subagents.

Verification: You pass chat_template_kwargs: {"verification": true}. A non-empty system message is required, otherwise you get an HTTP 400 error. The value must be a JSON boolean. The self-check appears in the reasoning_content field.

Recommended sampling parameters: temperature 1.0, top_p 0.95, top_k 20, min_p 0.0, presence_penalty 1.5.

What you need to try it

The source lists no minimum requirements. As a rule of thumb, a 36B model needs about 20–24 GB of memory at 4-bit quantization and considerably more (around 70 GB) at full precision. The examples run with vLLM on one GPU with 262K context. Long context costs extra memory (KV cache). A text-only mode (--language-model-only) saves memory because the image encoder is skipped. According to the source, there are no GGUF files, so running it with Ollama or on a Mac isn't directly supported right now. Without a big GPU, the model is unsuitable for Google Colab. Plan for about 70 GB of disk space for the weights.

Our take

Sage-A1.5 shows that open models in the 35B class are closing the gap on agent tasks against far larger systems, at least on the selected benchmarks. The subagent mode lifts BrowseComp from 81.7 to 82.8 and DeepResearch-Bench2 from 43.7 to 47.5, a concrete, measurable gain from orchestration. A note of caution: all figures come from the developers, partly with borrowed third-party values and inconsistent conditions. The benchmark selection is broad, but on several tests the big models are clearly ahead, and the predecessor sometimes does better. How well the model works day to day also depends heavily on your own harness, since you have to provide the orchestration yourself. For researchers and developers of agent systems, it's an interesting, freely usable starting point.