kkyyxhll/KAT-Coder-V2.5-Dev-fused-gate-up

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 27, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

KAT-Coder-V2.5-Dev is a 35.1 billion parameter Mixture-of-Experts (MoE) language model developed by kkyyxhll, with 3 billion activated parameters. Built upon Qwen3.6-35B-A3B, it is post-trained using SFT and RL to achieve state-of-the-art performance in Agentic Coding tasks among models of similar scale. This text-only model excels at code generation and problem-solving, supporting a context length of up to 262,144 tokens.

Loading preview...

KAT-Coder-V2.5-Dev: An Advanced Agentic Coding MoE Model

KAT-Coder-V2.5-Dev is a 35.1 billion parameter Mixture-of-Experts (MoE) model, with 3 billion activated parameters, developed by kkyyxhll. It is an open-weight release built on the Qwen3.6-35B-A3B base model, specifically designed for agentic coding tasks. This model has undergone extensive post-training, including Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), to enhance its performance and address abnormal behaviors.

Key Capabilities & Differentiators

  • State-of-the-Art Agentic Coding: Achieves leading performance in Agentic Coding benchmarks like SWE-bench Verified (69.40), SWE-bench Multilingual (63.00), and Terminal-Bench 2.1 (41.02), surpassing comparable models.
  • Optimized Behavior: Significant reduction in abnormal behaviors such as tool label errors (-9pp) and single-turn continuous repetition (-0.34pp) through targeted RL training.
  • Robust RL Training: Incorporates advanced RL techniques including Token-in-Token-out (TITO) consistency, Truncated Importance Sampling (TIS), reliable sandboxes, and hierarchical rewards based on execution feedback.
  • Context Length: Natively supports a context length of up to 262,144 tokens, with support for RoPE scaling techniques like YaRN for processing even longer texts.
  • Text-Only Model: This release focuses solely on language model weights; vision/multimodal components are not included.

Ideal Use Cases

  • Automated Code Generation and Repair: Excels in scenarios requiring autonomous code development and bug fixing.
  • Agentic Workflows: Suitable for building intelligent agents that interact with coding environments and tools.
  • Long-Context Code Analysis: Capable of processing and generating code within very large contexts, beneficial for complex projects.

Finetunes
1 models