Kimi-K3
Kimi K3 is an open-weight, 2.8T-parameter native multimodal agentic model with a 1M-token context window, designed for frontier coding, knowledge work, and reasoning tasks.
It addresses operational pain points in grok2api proxy egress by adding resilient failover/cooldown behavior and a heuristic quality breaker, while keeping changes auditable as a patch rather than a full fork.
这是一个非官方补丁分发仓库,为 chenyme/grok2api 增加固定代理快速恢复和可选的出口质量守护。仓库不复制上游完整源码,只发布可审计的 Git patch、功能说明和 AI 合并指南。
当前补丁基于:
v3.0.11090104504b403d65675a01dab9c92b3a235ee832MANIFEST.json 的 patch_commit(闭环 hard 隔离版)输出 Token / (总耗时 - 首字耗时),其中输出 Token 包含推理 Token。质量守护是启发式熔断器,不是模型能力鉴定器。中间层缓冲、已有文件、长常量或缓存内容可能造成异常高瞬时 Token/s。硬阈值策略偏激进,可按链路调高 hard_tps;软阈值仍以固定 Prompt 复测确认。
在干净的 grok2api 仓库中执行:
git fetch --tags origin
git checkout -b egress-enhancements v3.0.11
git am --3way /path/to/grok2api-egress-enhancements/patches/0001-feat-add-egress-recovery-and-quality-guard.patch
如果目标版本高于 v3.0.11,建议使用 AI 合并指南,让工具按功能不变量解决冲突,而不是直接覆盖新版文件。
go test ./...
python3 -m unittest -v tools/egress-quality-guard/quality_guard_test.py
cd frontend
pnpm lint
pnpm build
生产部署前还应验证:
transport error 冷却;补丁不包含部署配置。不要向 AI 工具提供真实 .env、config.yaml、数据库、状态卷、代理 URL、账号凭据或生产日志。合并时只需要上游源码、本仓库补丁和测试输出。
补丁在上游 MIT 许可框架下发布。保留上游项目的 LICENSE、版权信息和提交历史。本仓库不是 grok2api 官方发行版,也不代表上游维护者认可这些改动。
English documentation: README.en.md
Unofficial grok2api egress recovery and quality-guard patch kit
Similar projects matched by category, topics, and programming language.
Kimi K3 is an open-weight, 2.8T-parameter native multimodal agentic model with a 1M-token context window, designed for frontier coding, knowledge work, and reasoning tasks.
Harness Engineering is a methodology for improving coding agent outputs by carefully crafting the environment around them—providing curated context, tools, and executable constraints that encode an organization’s nonfunctional requirements and cumulative lessons.
A 28.9 million parameter language model runs on an $8 ESP32-S3 microcontroller entirely on-device, generating simple stories at about 9.5 tokens per second.