Ring-lite-2507: Boosted Deep Reasoning and Balanced General Capabilities

AntBailing's large model team announces Ring-lite-2507, a minor version upgrade of the Ring-lite series.

Deep Reasoning Capability Strengthened

Building on the previous C3PO‑based training that achieved stable Reasoning RL, the new version uses Ling‑lite‑base‑1.5 as the foundation and applies Long‑CoT SFT plus a Two‑staged RL pipeline. Additional general‑ability data and a new RL pipeline raise both reasoning and overall performance, with clear improvements on reasoning benchmarks such as ARC‑AGI‑v1.

Balanced and Comprehensive Ability

Coordinated training of general and reasoning skills leads to a more balanced model. The team notes that many open‑source reasoning models neglect general ability, but Ring-lite‑2507 demonstrates comparable general performance to qwen3‑8B. Issues like mixed Chinese‑English outputs and weak knowledge handling have been effectively resolved.

Training Process Upgrade

The team designed a Two‑staged RL pipeline. First, Long‑CoT SFT teaches the base model to think. Next, RLVR (a verifiable‑reward RL) enhances reasoning, followed by an RLHF stage to boost general ability. Experiments show that joint RLVR+RLHF training and the Two‑staged RL approach yield similar results, but due to differing difficulty levels, the Two‑staged RL scheme is preferred for engineering efficiency.

Model Open Source

Ring-lite-2507 shares the same base model as Ring-lite and can be used following the Ring-lite usage guide. The model is available for download on the project's GitHub repository, Hugging Face, and ModelScope. An upgraded version of C3PO is in development and will be released in the next version.

GitHub blog: https://inclusionai.github.io/zh/blog/ring-lite-2507

GitHub: https://github.com/inclusionAI/Ring

Hugging Face: https://huggingface.co/inclusionAI/Ring-lite-2507

ModelScope: https://www.modelscope.cn/models/inclusionAI/Ring-lite-2507