DATA CUT 2026-07-27 116 活跃模型 680 A 类案例

CASE EVIDENCE / A RECORD

loke-jad 使用 MiniMax-M2.5 处理软件工程任务执行

loke-jad 公开的代码代理与软件工程案例,来源为 公开代码库,复核于 2026-06-26T23:55:00Z。

原始记录:Claude Agent SDK + local MiniMax M2.5 via LiteLLM — MCP gateway coding agent with safety hooks

A

Chinese Brief

中文案例导读

loke-jad 公开的代码代理与软件工程案例,来源为 公开代码库,复核于 2026-06-26T23:55:00Z。 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。 Model Atlas 不把 benchmark、教程、发布说明或集合页包装成真实案例。

厂商MiniMax
模型MiniMax-M2.5
任务类型代码代理与软件工程
审核状态auto_approved

任务

真实任务背景

这是一个围绕软件工程任务执行的真实任务,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:Claude Agent SDK wired to a local llama-server running MiniMax-M2.5 (quantized, ~101GB) via LiteLLM on an NVIDIA DGX Spark (128GB unified memory). Includes MCP (Model Context Protocol) gateway for tool integration and s…

代码代理与软件工程公开代码库A 类可核验real_case
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:Working local coding agent with full Claude Code compatibility. Detailed optimization plan documented (OPTIMIZATION.md): expected 10-20x prompt processing speedup via persistent KV cache warming. Agent runs fully locall…

模型作用

MiniMax-M2.5 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:MiniMax M2.5 (via Unsloth GGUF quantization, llama-server, and LiteLLM proxy) serves as the sole LLM backbone for the coding agent. The model handles code generation, tool orchestration, multi-turn conversation, and rea…

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:Project marked as 'Dormant' by author. No README.md file — evidence comes from OPTIMIZATION.md and GitHub page metadata. 0 stars. Strong technical evidence of real deployment (detailed VRAM measurements, cache profiling…