Sage-A1.5

Shanghai AI Laboratory's 35B-A3B agent model on Qwen3.6, with parallel sub-agents and a self-review mode. Apache 2.0.

In short. Sage-A1.5 is an open AI model built to work as a research and coding agent. It plans a task, can hand parts of it to helper agents that work in parallel, and can stop to check its own progress before it continues. It reads images as well as text, and a 4-billion-parameter sibling, Agents-A1-4B, runs on much smaller machines.

Summary

Sage-A1.5 comes from InternScience, the open-source group of the AI for Science Center at Shanghai AI Laboratory. It is a 35B mixture-of-experts agent model with about 3B parameters active per token, built on Alibaba's Qwen3.6-35B-A3B and on the lab's earlier Agents-A1. It takes text and images, thinks in a reasoning block, and has a native 256K-token context that the card extends to 512K with YaRN. Its repository appeared on Hugging Face on 15 October 2026 and the technical report is dated 20 October. Agents-A1 came out in June 2026 after a three-stage recipe of full-domain supervised fine-tuning, domain-level teacher models and multi-teacher on-policy distillation, and the lab added a dense Agents-A1-4B model in July.

Sage-A1.5 adds two opt-in modes. In agentic orchestration, you register a spawn_subagent tool and the model fans independent sub-questions out to parallel sub-agents that each have their own context. Verification mode is switched on with a chat-template flag and makes the model pause mid-reasoning to review what it has established and what has failed. On its card, the lab reports 95.1 on GAIA, 81.7 on BrowseComp (82.8 with orchestration), 61.2 on SWE-bench Pro and 92.0 on XBench-DS-2510. For SWE-bench Pro the table shows 49.5 for Qwen3.6-35B-A3B, 59.6 for Ornith-1.5 and 64.6 for GPT-5.6-sol. On BrowseComp it shows 67.9 for Qwen3.6-35B-A3B and 91.2 for Kimi-K3, and Agents-A1 still scores higher on GAIA, 96.0.

All figures are the lab's own, including its runs of the competing models. The card gives a single-GPU vLLM recipe at 262K context, a text-only mode that skips the vision encoder to free KV-cache memory, and recommended sampling of temperature 1.0, top-p 0.95 and top-k 20. The bf16 weights are about 71.9 GB, and the official Q8_0 and Q4_K_M GGUF builds are 37.8 GB and 21.7 GB, each with a 0.9 GB vision projector. For the 4.54B dense Agents-A1-4B, the lab reports 66.8 on BrowseComp and 94.8 on IFEval, and its Q4_K_M GGUF is 2.71 GB. In the 30 days to 11 November 2026, Hugging Face counted about 620,000 downloads for the Sage-A1.5 Q4_K_M GGUF, 528,000 for the Q8_0 GGUF, 21,000 for the bf16 repository and 12,600 for Agents-A1.

Good for

  • Research and coding agents that need long-horizon search, tool calling and up to 256K tokens of context
  • Agent harnesses that can run parallel sub-agents through a spawn_subagent tool
  • Running a 35B-A3B agent locally with the official Q4_K_M GGUF of about 22 GB, or a small agent with the 4B model
  • Teams that want an Apache 2.0 agent model with image input on a known Qwen3.6 base

Watch out for

  • Every benchmark number comes from InternScience, including its runs of competing models, and the card's own table shows larger models such as Kimi-K3 and GPT-5.6-sol ahead on several rows
  • Agentic orchestration needs your own harness, because the model only calls the tool and your code must run the sub-tasks, with batch size and turn limits set on the harness side
  • Verification mode needs a non-empty system message and a JSON boolean flag in the chat template kwargs, or the request is rejected with HTTP 400
  • Agents-A1 and Agents-A1-4B are earlier models with their own benchmark tables, so check which repository you download

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the Agents-A1-4B (Q4_K_M GGUF) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
Agents-A1-4B BF16 4.54B 10 GB 5.22 GB 3.13 GB Laptop or phone
Agents-A1-4B (Q4_K_M GGUF) INT4 4.54B 10 GB 5.22 GB 2.98 GB Laptop or phone
Agents-A1 BF16 35.1B 3B active 81 GB 40 GB 24 GB Gaming GPU
Sage-A1.5 BF16 36B 3B active 83 GB 41 GB 25 GB Gaming GPU
Sage-A1.5 (Q8_0 GGUF) INT8 36B 3B active 83 GB 41 GB 25 GB Gaming GPU
Sage-A1.5 (Q4_K_M GGUF) INT4 36B 3B active 83 GB 41 GB 24 GB Gaming GPU

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Sage-A1.5?
Sage-A1.5 is an open AI model built to work as a research and coding agent. It plans a task, can hand parts of it to helper agents that work in parallel, and can stop to check its own progress before it continues. It reads images as well as text, and a 4-billion-parameter sibling, Agents-A1-4B, runs on much smaller machines.
Can I use Sage-A1.5 commercially?
Yes. Sage-A1.5 is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Sage-A1.5 open source?
Sage-A1.5 is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Sage-A1.5 on my own computer?
Yes. The smallest variant, Agents-A1-4B (Q4_K_M GGUF), needs about 2.98 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Sage-A1.5 released?
Sage-A1.5 was released by Shanghai AI Laboratory on October 15, 2026.
How long a context window does Sage-A1.5 have?
Sage-A1.5 supports a context window of up to 256K tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Qwen3.6 Alibaba Qwen 27.8B – 36B 3B active 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Ornith 1.5 DeepReinforce 9.65B – 403B 17B active 256K MIT (Permissive) 6.66 GB Laptop or phone
Qwen3.8-27B Alibaba Qwen 27.8B 256K Apache 2.0 (Permissive) 19 GB Gaming GPU
Kimi K3 Moonshot AI 2.8T 104B active 1M Kimi K3 License (Conditional) 1.72 TB Multi-node cluster