GitHub radar
Sage-A1.5: a local agent model for coding and tasks
Shanghai AI Lab released Sage-A1.5, an agent model for coding and tasks built on their own Qwen3.6, which a person can install and run locally.
Sage-A1.5 is a mixture-of-experts model with 35 billion parameters, which Shanghai AI Lab built on their own Qwen3.6 and on their earlier model Agents-A1. It can break a long task into subtasks on its own, call tools, and check its own answer before giving it out, and the earlier version, A1, did not do this. Besides coding, the model is aimed at scientific and engineering tasks: multi-step problem solving, work with scientific literature, function calls. Compared to the previous version it gained on the SWE-bench Pro and MLE-Lite tests, and on the GAIA test it scored 95.1, which, according to the developers, is higher than the result of models larger than itself. The model is distributed under the Apache 2.0 license and runs through standard engines such as vLLM, SGLang, or Transformers, and there is no ready Ollama build yet.
Why a vibe-coder should care
For coding and agentic tasks this is one of the few open models that a person can install locally and give real multi-step tasks, not just chat. But the benchmark numbers are one thing, and real agentic work day after day is another, and the test does not show this. I would use such a model not instead of Opus or Fable in Claude Code, but for simple repeated tasks or when there is no paid subscription at hand.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install the Sage-A1.5 model by InternScience for me: https://huggingface.co/InternScience/Sage-A1.5 — pick the right version for my hardware, run it through vLLM or Transformers, give it a simple work task, and show me the result.
You need a Mac with 64 GB of memory or more, or a GPU with 48 GB of VRAM — in the 4-bit version the model takes up about 22 GB.
Open on Hugging Face▌ More finds