InternScience lists Sage-A1.5 on Hugging Face as an agentic ~35B Mixture-of-Experts model built on Qwen3.6 and Agents-A1 from Shanghai AI Laboratory. Weights are Apache 2.0. The same card points to ModelScope, an FP8 sibling (InternScience/Sage-A1.5-FP8), and GitHub InternScience/Agents-A1 for evaluation code. The technical report’s arXiv stamp is UTC 2026-10-20 17:32:48—about Beijing time 2026-10-21 01:32.

Two advanced modes ship off by default. Verification Mode is opt-in self-review inside the model’s own chain of thought: enable it per request with `chat_template_kwargs` set to `verification: true` (JSON boolean, not a string). A non-empty system message is required. Agentic Orchestration is enabled by registering a `spawn_subagent` tool; your deployment must actually run those sub-tasks and return structured results—the weight file alone does not spawn workers. Native context is 256K tokens; extending to 512K needs YaRN (`factor: 2.0`) in the serving config. Serving recipes on the card cover SGLang and vLLM.

The model card and paper report strong scores on agent and science benchmarks against other ~35B and larger systems. Those figures are InternScience’s own reported numbers. Related reading: MiniMax M3.1 Flash Open Weights Are on Hugging Face, Kimi Adds K3.1 Pro, and What Ornith-1.5-122B Needs to Run and How to Start It With Ollama.

For how Sage-A1.5 actually performs on your workloads, how much VRAM it needs, and whether Verification or Orchestration help in practice, we have not tested yet. If you have used it, you can reach the site owner through the contact link in the footer. The Hugging Face card also links a Free Trial portal on InternScience’s discovery site; that is separate from downloading open weights, and we have not used that portal either.

Sources

  • Hugging Face: InternScience/Sage-A1.5 — https://huggingface.co/InternScience/Sage-A1.5
  • Hugging Face: InternScience/Sage-A1.5-FP8 — https://huggingface.co/InternScience/Sage-A1.5-FP8
  • ModelScope: InternScience/Sage-A1.5 — https://modelscope.cn/models/InternScience/Sage-A1.5
  • GitHub: InternScience/Agents-A1 — https://github.com/InternScience/Agents-A1
  • arXiv: Sage-A1.5: From Long-Horizon Execution to Scientific Exploration with Capability Composition — https://arxiv.org/abs/2610.28578
  • InternScience Free Trial portal — https://discovery.intern-ai.org.cn/token-plan/home?tabIndex=0&lang=en-US
  • Project homepage: Sage-A1.5 — https://alpha-innovator.github.io/Sage-A1.5/

Recommended reads