Kimi K2.5

sangjiexun 公开的研究与报告生成案例，来源为 公开代码库，复核于 2026-06-26T23:14:00Z。 Model Atlas 将它标记为 A 类证据，因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。

任务：这是一个围绕研究分析和报告生成的真实任务，公开材料可以回溯到具体使用者和具体产物。 原始资料写作：Developer sangjiexun built ClawCoder, a pip-installable interactive CLI coding assistant powered by Kimi K2.5 through the Moonshot AI API (api.kimi.com/coding, model: k2p5). The tool provides AI-powered conversation, fi…

公开产物：公开材料提供公开代码、README 或项目配置，可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作：A fully functional CLI tool published on PyPI (pip install clawcoder) with features including REPL mode, directory-aware file operations, AI-driven code editing, binary disassembly, and persistent chat history. The tool…

模型作用：Kimi K2.5 在该案例中承担研究分析和报告生成相关的生成、分析、编排或实现角色。 原始资料写作：Kimi K2.5 (via model identifier k2p5) is the sole AI backend powering all intelligent features: natural language conversation, code generation and editing, file analysis, and multi-modal understanding. The 256K context …

A 类理由：有具体使用者、具体任务、公开原始证据和可访问产物，可公开核验。

风险边界：当前判断基于公开材料；若产物下线、仓库变更或模型参与比例仅来自作者自述，需要在引用前重新复核。 原始资料写作：Small open-source project (7 stars). Developer is an individual contributor. Tool uses Kimi K2.5's proprietary API endpoint.

原始记录：sangjiexun builds ClawCoder CLI coding assistant powered by Kimi K2.5 via LangGraph

GuanxingLu 公开的研究与报告生成案例，来源为 公开代码库，复核于 2026-06-26T23:14:00Z。 Model Atlas 将它标记为 A 类证据，因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。

任务：这是一个围绕研究分析和报告生成的真实任务，公开材料可以回溯到具体使用者和具体产物。 原始资料写作：Researcher GuanxingLu built OpenPARL, a from-scratch reproduction of Kimi K2.5's PARL (Parallel Agent Reasoning Layer) Agent Swarm as described in the paper arXiv:2602.02276 'Kimi K2.5: Visual Agentic Intelligence'. The…

公开产物：公开材料提供公开代码、README 或项目配置，可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作：A complete open-source reproduction of the K2.5 PARL agent swarm architecture, including RL training pipelines, evaluation scripts, and ablation studies. The project demonstrates that K2.5's PARL agent swarm methodology…

模型作用：Kimi K2.5 在该案例中承担研究分析和报告生成相关的生成、分析、编排或实现角色。 原始资料写作：Kimi K2.5 provides the foundational research paper and architecture specification (arXiv:2602.02276) that OpenPARL reproduces. The PARL Agent Swarm design pattern from K2.5 is the core innovation being studied, includin…

A 类理由：有具体使用者、具体任务、公开原始证据和可访问产物，可公开核验。

风险边界：当前判断基于公开材料；若产物下线、仓库变更或模型参与比例仅来自作者自述，需要在引用前重新复核。 原始资料写作：This is a research reproduction, not direct K2.5 API usage. The actual models used are Qwen3-4B variants. The value is in validating and open-sourcing K2.5's architectural innovations for the broader research community.

原始记录：GuanxingLu reproduces Kimi K2.5 PARL Agent Swarm architecture on WideSearch with RL-trained orchestrator

阿布AI进化论 (Bilibili creator, mid:499994152) 公开的多模态生成与理解案例，来源为 bilibili，复核于 2026-06-26T23:14:00Z。 Model Atlas 将它标记为 A 类证据，因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。

任务：Bilibili creator 阿布AI进化论 built and deployed a production AI assistant system by integrating OpenClaw with Kimi K2.5 as the underlying model and Feishu (Lark) bot as the communication interface, creating a 7x24 always-on AI assistant for work. The tutorial demonstrates the complete Windows-based setup from OpenClaw installation through Kimi K2.5 model configuration to Feishu bot webhook deployment.

公开产物：公开材料提供原始证据链接和可访问产物，可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作：A fully functional 7x24 enterprise AI assistant deployed via Feishu bot, powered by Kimi K2.5 through OpenClaw. The video received 25,575 views and 384 likes. Published on 2026-02-11, indicating early-adopter deployment…

模型作用：Kimi K2.5 在该案例中承担多模态内容处理相关的生成、分析、编排或实现角色。 原始资料写作：Kimi K2.5 is the core reasoning model powering the AI assistant's conversational intelligence, task execution, and multi-turn dialogue capabilities through the OpenClaw framework. The model handles natural language unde…

A 类理由：有具体使用者、具体任务、公开原始证据和可访问产物，可公开核验。

风险边界：当前判断基于公开材料；若产物下线、仓库变更或模型参与比例仅来自作者自述，需要在引用前重新复核。 原始资料写作：Tutorial-format video. While it demonstrates real deployment, the primary artifact is a setup guide rather than a production usage report. Published by a single content creator.

原始记录：阿布AI进化论 deploys OpenClaw + Kimi K2.5 + Feishu bot as 7x24 enterprise AI assistant

鲲鹏Talk (Bilibili creator, mid:18052876) 公开的代码代理与软件工程案例，来源为 bilibili，复核于 2026-06-26T23:14:00Z。 Model Atlas 将它标记为 A 类证据，因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。

任务：Bilibili tech creator 鲲鹏Talk demonstrated end-to-end development of a fund company portal website using Kimi Code paired with Kimi K2.5 model. The video shows the complete development workflow as a practical replacement for Claude Code, covering frontend layout, backend logic, and full-stack integration. The creator explicitly positioned Kimi K2.5 as a viable alternative to Claude Code for real software development projects.

公开产物：公开材料提供原始证据链接和可访问产物，可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作：Successfully built and demonstrated a complete fund company portal website with Kimi Code + K2.5. The video received 64,063 views and 306 likes, indicating the demo was well-received by the developer community. Publishe…

模型作用：Kimi K2.5 在该案例中承担软件工程任务执行相关的生成、分析、编排或实现角色。 原始资料写作：Kimi K2.5 serves as the core coding intelligence behind Kimi Code, handling all code generation, debugging, and project scaffolding tasks throughout the website development process. The model's agentic coding capabiliti…

A 类理由：有具体使用者、具体任务、公开原始证据和可访问产物，可公开核验。

风险边界：当前判断基于公开材料；若产物下线、仓库变更或模型参与比例仅来自作者自述，需要在引用前重新复核。 原始资料写作：Video demonstration by a single creator. The website is demonstrated in the video but the actual deployed site URL is not provided. Bilibili video is the primary evidence.

原始记录：鲲鹏Talk develops fund company portal website using Kimi Code + K2.5 as Claude Code alternative

Moonshot AI Research Team 公开的代码审查与测试案例，来源为 官方页面，复核于 2026-06-26T22:55:00Z。 Model Atlas 将它标记为 A 类证据，因为它同时具备具体使用者、具体任务、公开原始证据和可访问产物。

任务：这是一个围绕代码审查和测试生成的真实任务，公开材料可以回溯到具体使用者和具体产物。 原始资料写作：Evaluate Kimi K2.5 on WorldVQA, a novel benchmark testing atomic visual knowledge of frontier models across 9 task categories (People, Objects, Culture, Geography, Transportation, Brands, Sports, Entertainment, Location…

公开产物：公开材料提供原始证据链接和可访问产物，可用于核验任务结果、项目形态和模型绑定关系。 原始资料写作：Kimi K2.5 achieved 46.3% accuracy on WorldVQA Overall Accuracy, ranking second only to Gemini-3-pro (47.4%) and outperforming Claude-opus-4.5 (36.8%), GPT-5.2 (28.0%), Qwen3-VL-235B (23.5%), and all other evaluated mode…

模型作用：Kimi K2.5 在该案例中承担代码审查和测试生成相关的生成、分析、编排或实现角色。 原始资料写作：Kimi K2.5 demonstrates state-of-the-art visual knowledge understanding, achieving near-Gemini-3-pro performance on a challenging benchmark that tests real-world visual knowledge rather than OCR or simple recognition. Th…

A 类理由：有具体使用者、具体任务、公开原始证据和可访问产物，可公开核验。

风险边界：当前判断基于公开材料；若产物下线、仓库变更或模型参与比例仅来自作者自述，需要在引用前重新复核。 原始资料写作：Published on Moonshot AI's official blog with arxiv paper reference. Benchmark data is from the model provider's own research but with open-source reproducible evaluation.

原始记录：K2.5 Powering WorldVQA Visual Knowledge Benchmark Research