Moonshot AI's K1.5, released before K2, has a core focus on using pure reinforcement learning to improve reasoning capabilities.
Technical approach
K1.5's reasoning enhancement does not rely on the conventional path of SFT (supervised fine-tuning) or RLHF. Instead, it depends more heavily on pure RL (reinforcement learning) training.
This is similar to the approach behind DeepSeek R1 — letting the model learn to reason through extensive trial and error and reward signals, rather than being taught with human-annotated "correct reasoning steps."
The benefits are:
- Not limited by the quality or quantity of human-annotated data
- The model may discover reasoning strategies that human annotators never considered
- Theoretically easier to scale up
Performance
K1.5 achieved top-tier results among open-source models of its generation on mathematical reasoning benchmarks. While there is still a gap compared to closed-source models like GPT-4o and Claude Opus, the gap is within an acceptable range.
More importantly, K1.5 validated the feasibility of the pure RL approach on a medium-scale model. This laid the technical foundation for the subsequent development of K2 and K2.5.
Relationship with K2
K1.5 can be seen as a technical preview version for K2 — after validating the RL-driven reasoning approach, K2 directly applied this methodology at a trillion-parameter scale, while also incorporating a MoE architecture to control computational costs.
The iterative process from K1.5 to K2 shows that Moonshot AI's R&D strategy is to validate the methodology first, then scale up, rather than starting with a massive parameter count from the beginning. This gradual approach is pragmatic for a team with limited resources.
Sources: CocoLoop, Moonshot AI technical blog