Neura Market
Kimi k1.5: Scaling Reinforcement Learning with LLMs logo

Kimi k1.5: Scaling Reinforcement Learning with LLMs

Free

Kimi k1.5 — an o1-level multi-modal model

FreeFree tier
Inputs: textOutputs: text
Type
Open Source
Company
MoonshotAI
Visit Website GitHub (40 stars)

About Kimi k1.5: Scaling Reinforcement Learning with LLMs

Kimi k1.5 is a multi-modal language model developed by MoonshotAI that achieves o1-level reasoning performance through scaling reinforcement learning with long context windows (128k tokens). It demonstrates state-of-the-art short-CoT results on math and coding benchmarks (AIME, MATH-500, LiveCodeBench) and matches o1's long-CoT performance across multiple modalities. Key innovations include long context scaling via partial rollouts for efficient training, improved policy optimization using online mirror descent, and a simplistic framework that avoids complex techniques like Monte Carlo tree search and process reward models. The model is jointly trained on text and vision data, enabling reasoning over both modalities.

Key Features

Long context scaling up to 128k tokens
Improved policy optimization with online mirror descent
Simplistic RL framework without Monte Carlo tree search or value functions
Multi-modal reasoning on text and vision data
State-of-the-art short-CoT performance on math and coding benchmarks
Long-CoT performance matching o1 across modalities

Pros & Cons

Pros
  • Outperforms GPT-4o and Claude Sonnet 3.5 on short-CoT tasks by large margins
  • Matches o1's long-CoT performance across multiple modalities
  • Efficient training via partial rollouts reduces computational cost
  • Simplistic framework avoids complex techniques
  • Open source and publicly available on GitHub
Cons
  • No released model weights or executable version available yet
  • Requires large context window (128k) which may be computationally intensive
  • Limited documentation beyond research paper and README

Best For

Mathematical reasoning (AIME, MATH-500)Coding and programming challenges (LiveCodeBench, Codeforces)Multi-modal reasoning tasks (MathVista)General reinforcement learning research with LLMs

FAQ

What is Kimi k1.5?
Kimi k1.5 is a multi-modal language model developed by MoonshotAI that achieves o1-level reasoning through scaling reinforcement learning with LLMs.
What makes Kimi k1.5 different from other models?
It uses long context scaling up to 128k tokens, a simplistic RL framework without Monte Carlo tree search or process reward models, and is jointly trained on text and vision data.
What benchmarks does Kimi k1.5 excel at?
It achieves state-of-the-art short-CoT performance on AIME, MATH-500, and LiveCodeBench, and matches o1 on long-CoT tasks across MathVista, AIME, and Codeforces.
Is Kimi k1.5 open source?
Yes, the paper and code are available on GitHub under MoonshotAI/Kimi-k1.5.