CPU3DAI tools for 3D, video and audio

InternScience's Sage-A1.5: a new 36B-parameter agentic model

The model InternScience/Sage-A1.5 with the text-generation task has been published on HuggingFace. It is an agentic model on a mixture-of-experts architecture, built on the basis of Qwen3.6 and Agents-A1 from Shanghai AI Laboratory. The author highlights three key capabilities: agent orchestration with parallel subtasks, verification of long reasoning chains, and open scientific research.

What it means

The model contains 36 billion parameters and is distributed under the apache-2.0 license. The full weights in the repository take up 67 GB in safetensors format, with the largest file at 5 GB. It runs via transformers, and compatibility with vLLM and SGLang is also claimed. For local running, the site has 13 quantized builds: GGUF, FP8 and MLX. The most compact working version is IQ4_XS by na0x2c6a, which weighs 18 GB. The developer does not specify VRAM requirements. The model is aimed at long scientific tasks with decomposition and verification of intermediate steps. You can follow the emergence of similar tools in the CPU3D news feed.
Sage-A1.5 by InternScience agentic model with 36B parameters ? Scientific task long, complex needs decomposition Sage-A1.5 36B parameters mixture-of-experts base: Qwen3.6 + Agents-A1 Agent orchestration parallel subtasks yes Reasoning verification long step chains Open research scientific tasks Launch transformers, vLLM, SGLang 13 quantized builds GGUF, FP8, MLX Weights 67 GB safetensors max file — 5 GB IQ4_XS by na0x2c6a 18 GB, compact License apache-2.0
How the method works. The diagram was drawn based on this news note.

See also