Kimi k1.5 — January 2025

What Changed

Moonshot AI's Kimi k1.5 demonstrated that long-context scaling — extending the policy's context window to 128K tokens — is itself a key axis for improving reasoning quality. Without MCTS, value functions, or process reward models, Kimi k1.5 matched OpenAI o1 on AIME (77.5) and MATH 500 (96.2) using a simpler RL framework.

The insight: a long context window lets the model plan, reflect, and self-correct within a single forward trajectory, because earlier reasoning steps remain in-context and can be referenced and revised.

Key Technical Details

Partial rollout reuse avoids regenerating entire trajectories from scratch. When exploring long reasoning chains, the system reuses large chunks of already-computed trajectories and only generates novel suffixes. This dramatically reduces the per-training-step compute cost for 128K-token sequences.

Online mirror descent replaces standard gradient descent for policy optimization in this long-context regime:

\[ \pi_{t+1} = \arg\min_\pi \left[ -\mathbb{E}_\pi[R] + \frac{1}{\eta} D_{\mathrm{KL}}(\pi \| \pi_t) \right] \]

In Plain English

Mirror descent keeps each policy update close to the previous policy (via a KL ball) while maximizing expected reward. This is more stable than unconstrained gradient steps when the reward landscape is non-smooth over long sequences.

Length penalty: a soft penalty discourages unnecessarily long chains while preserving chains that need depth for correctness.

Technical Details

• No MCTS: MCTS requires many forward passes and a value function. Kimi k1.5 shows that a plain RL loop with long context achieves comparable benchmark results — a simpler, more scalable alternative.
• Short-CoT regime: Kimi k1.5 also achieves strong results in the short-CoT regime (no explicit reasoning steps in the prompt), outperforming GPT-4o and Claude Sonnet 3.5 by up to +550% on some tasks.
• Multimodal: trained jointly on text and vision data, enabling reasoning over image inputs.

Practical Implications

Long-context RL is now a viable alternative to search-augmented inference. When deploying reasoning models, measure cost at the reasoning-token level, not just final output tokens — the 128K traces are the expensive part.

Interview Questions

1. Why does extending a model's context window improve its reasoning ability in RL training?
2. How does online mirror descent stabilize policy updates compared to vanilla gradient ascent on reward?
3. Kimi k1.5 achieves o1-level results without MCTS — what does this tell us about the relationship between search and in-context reasoning?

Code Example

(No standalone code sample in the source entry; partial rollout reuse and online mirror descent are primarily training-stack concerns.)