Home › Language Models › Kimi-Dev

Kimi-Dev - AI Language Models Tool

Overview

Kimi-Dev (Kimi-Dev-72B) is an open-source coding LLM built for automated software engineering workflows: automated bug fixing, generation of unit tests, and end-to-end repository patching that verifies fixes by executing real test suites. The model was developed by Moonshot AI and is presented as a duo of specialized behaviors — BugFixer and TestWriter — that first localize relevant files and then produce concrete edits or tests; training emphasizes producing patches that make the entire test-suite pass before accepting changes. According to the project materials, Kimi-Dev-72B achieves 60.4% on SWE-bench Verified among open-source models, and its training pipeline includes a mid-training stage (~150B tokens) built on a Qwen-2.5-72B base followed by a large-scale reinforcement learning phase that uses outcome-only rewards based on Dockerized test execution. ([moonshotai.github.io](https://moonshotai.github.io/Kimi-Dev/?utm_source=openai)) The release is intended for researchers and engineers who want reproducible, verifiable code-repair workflows: model weights and code are available on GitHub and Hugging Face, and the repository provides deployment guidance (vLLM serving, CUDA 12.8 recommendations) plus example rollout scripts for BugFixer/TestWriter. Community reports highlight strong debugging performance in many real-world scenarios, while some users note inference speed and hardware requirements as practical constraints for local deployments. The project is positioned as open-access research code with an accompanying technical report and demo resources. ([github.com](https://github.com/MoonshotAI/Kimi-Dev))

GitHub Statistics

  • Stars: 1,087
  • Forks: 130
  • Contributors: 3
  • License: NOASSERTION
  • Primary Language: Python
  • Last Updated: 2025-09-30T02:15:54Z

The GitHub repository shows a high level of interest (≈1.1k stars, 130 forks) while active maintenance and community contribution are relatively limited compared to those counts: the repository lists a small set of contributors and only a few open issues at the time of inspection. The README contains detailed installation and usage instructions (conda environment, pip install -e ., vLLM serving examples) and links to model weights on Hugging Face, but the codebase surface is compact (Python-only modules) and the contributor count suggests early-stage community development rather than a broad contributor ecosystem. These indicators imply strong adoption interest but a smaller maintainer base, so organizations planning production integration should evaluate long-term support and perform their own testing. ([github.com](https://github.com/MoonshotAI/Kimi-Dev))

Installation

Install via pip:

git clone https://github.com/MoonshotAI/Kimi-Dev.git
conda create -n kimidev python=3.12
conda activate kimidev
pip install -e .
# Install vLLM with CUDA 12.8 (recommended for serving)
pip install vllm --extra-index-url https://download.pytorch.org/whl/cu128
# Serve the model with vLLM (example)
vllm serve Kimi-Dev-72B --served-model-name kimi-dev --host 0.0.0.0 --port 8000 --gpu-memory-utilization 0.95 --max-seq-len-to-capture 131072 --tensor-parallel-size 8

Key Features

  • Kimi-Dev-72B: a 72B-parameter coding LLM optimized for bug fixes and test generation.
  • Duo workflow (BugFixer + TestWriter) — file localization then targeted code edits.
  • Reinforcement learning with outcome-only rewards: accepts patches only when test suites pass.
  • Mid-training on ~150B curated tokens drawn from GitHub issues and PR commits.
  • Test-time self-play and scaling: generates up to 40 patch and 40 test candidates per issue.
  • Deployment-ready guidance (vLLM serving, CUDA 12.8) and downloadable weights on Hugging Face.

Community

The project has attracted significant attention (≈1.1k GitHub stars and 130 forks) and active discussion in developer communities. Users on Reddit and technical blogs report strong debugging and test-generation capabilities but note higher hardware needs and slower local inference in some configurations. Moonshot AI provides demos, a Hugging Face model card, and a technical blog describing training details; third-party coverage (industry press and community posts) highlights the model's SWE-bench results and practical promise while urging independent validation for production use. Contributors and maintainers are currently few, so expect community growth to be incremental and evaluate operational support before large-scale integration. ([github.com](https://github.com/MoonshotAI/Kimi-Dev))

Last Refreshed: 2026-01-09

Key Information

  • Category: Language Models
  • Type: AI Language Models Tool
Visit Official Website

Related Tools

Qwen2.5-7B

Qwen2.5-7B is a large language model designed for text generation, featuring improvements in coding, mathematics, instruction following, long text generation, and multilingual support. It supports context lengths up to 128K tokens and is intended for sophisticated NLP tasks.
DeepSeek-V3

Open‑weight flagship LLM (base and chat) with strong reasoning and tool‑use; runs with SGLang/LMDeploy/vLLM and permissive model license.
Llama 3

Llama 3 is an open access large language model (LLM) released by Meta, available in various configurations (8B and 70B parameters) with capabilities for fine-tuning and integrations into platforms like Hugging Face, Google Cloud, and Amazon SageMaker.
UNfilteredAI-1B

A large-scale text generation model designed for creative and unconstrained content production without traditional filtering.
Shuttle-3

Shuttle-3 is a state-of-the-art language model designed for high-quality text generation, particularly suited for complex chat, multilingual communication, and reasoning tasks. It is fine-tuned from the Qwen-2.5-72b-Instruct model and designed to emulate high-quality prose similar to Claude 3 models.
WizardLM

WizardLM is a state-of-the-art large language model designed for complex chat, multilingual tasks, reasoning, and agent functionalities. It features an AI-powered pipeline (Auto Evol-Instruct) that optimizes instruction datasets for improved performance across various domains and leverages Arena Learning for an expanded learning pool of challenging instruction data.