CogEvol-4B-GGUF

CogEvol-4B is an open, post-trained model built on Qwen3.5-4B for Learning Environment Generation (LEG) — given a natural-language course brief, it generates complete, usable learning artifacts in a single pass, either as a structured-JSON slide page or a self-contained interactive HTML page (simulations, visualizations, interactive exercises) that runs directly in the browser. As the small open member of the CogEvol family, it was trained through a three-stage recipe — mixed SFT, followed by Slide RL, then interactive-HTML RL — on 53,687 verified samples, using a hybrid rule-and-VLM reward system hardened against reward hacking so that interactivity is measured via automated probes rather than judged superficially, as detailed in the accompanying paper "CogEvol: Towards Efficient and Reliable Learning Environment Generation" (arXiv:2608.30968). It's deployable via SGLang, vLLM, or llama.cpp (with a quantized Q4_K_M GGUF at ~2.4GB also available), requires thinking mode disabled at inference and the shipped system prompt template for slide generation, and integrates with the OpenMAIC app for fully offline use, with a complete deployment guide, serving scripts, and evaluation tools available on the project's GitHub repository; it's released under the Apache 2.0 license.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp