Klear

model paper
Kuaishou
Sparse MoE LLM family from Kuaishou, including a 46B-parameter base model (2.5B active) and the Klear-Reasoner variant achieving 90.5% on AIME 2024 via Gradient-Preserving Clipping Policy Optimization (GPPO).
HuggingFaceGitHubPaper (arXiv)

Outputs 3

Klear-46B-A2.5B

model

Sparse MoE LLM with 46B total parameters and 2.5B active, designed for high performance and inference efficiency.

HuggingFace
Architecture MOE
Parameters 46B
Active params 2.5B

Klear-Reasoner: Advancing Reasoning Capability via GPPO

paper

Technical report on Gradient-Preserving Clipping Policy Optimization (GPPO). Based on Qwen3-8B-Base, achieves 90.5% on AIME 2024 and 83.2% on AIME 2025.

Paper (arXiv)GitHubHuggingFace

arXiv: 2508.07629

Klear-Reasoner-8B

model

8B reasoning model based on Qwen3-8B-Base, fine-tuned with long CoT SFT and RL on math and coding tasks. 90.5% on AIME 2024, 83.2% on AIME 2025.

HuggingFaceGitHub
Architecture DENSE
Parameters 8B
moereasoningopen-weightcoding