NVIDIA Nemotron 3.5 Lightning：Kilo实测与Agent工作流整合
NVIDIA发布Nemotron 3.5 Lightning开源模型，30B MoE架构激活3B参数，支持100万token上下文，推理速度号称比竞品快4倍。Kilo实测认为适合常驻Agent工作流，模型性格鲜明会主动表达观点。
评论
点赞
收藏Metabase Incident Impacting Kilo Code Customer Data
August 9, 2026 14:50 UTC Update We continue to investigate the impact to Kilo users from the Metabase incident. Our investigation so far has confirmed that the Kilo Slackbot was impacted and a small s...
Meta Will Cut Your Token Bill 20x. The Price Is Your Prompts.
Meta shipped its first coding agent on August 5. Muse Code installs from the terminal with one command, runs on macOS and Linux, and is powered by Muse Spark 1.2, a new coding-focused model. It plans ...
How Many Spoons Does Your AI Coding Tool Cost?
Spoon theory is a way of describing a limited daily budget of energy, coined by people managing chronic illness to explain why a task that looks small can still cost everything they have left. Develop...
Your Agent Has Too Much Context
There’s a reflex that every team adopting coding agents develops: when the agent gets something wrong, add more context. Write a spec. Add a rule to the instructions file. Tell it to go read the relev...
The Subtle Art of Tokenrelaxxing
At the Stanford AI Summit last weekend, Arvind KC, Chief People Officer at OpenAI, asked the audience a question: How many of you find that AI has made you more productive? In response, almost every h...
Kilo Code’s JetBrains plugin was built for split mode from day one
The Kilo Code plugin for JetBrains IDEs is native. Not a webview, not a VS Code extension in a JetBrains frame. It’s built on the IntelliJ platform, and specifically on JetBrains’ split mode API, whic...
Code Review Has Changed. The Contract Needs to Change Too.
For years, code review has been one of the core rituals of software engineering. A developer writes code and opens a pull request. Another developer reads the diff, looks for bugs, checks the tests, q...
We analyzed 10,643 AI code reviews.
In late July, more than 230 organizations signed the Open Weights and American AI Leadership letter, asking Washington to protect the open-weight ecosystem rather than restrict it. Anaconda, our paren...
代码很便宜，而评审却很昂贵。
对于越来越多的贡献而言，如今从头重新生成拉取请求，比审核别人发给我们的拉取请求要容易得多。 这话听起来或许有些奇怪，但它真实地揭示了在编码代理时代，开源领域正在发生的一系列变化，并迫使我们重新思考那些已经习以为常、被我们视作理所当然的流程，而这些流程其实已经沿用了整整十年。 感谢您的阅读！欢迎订阅Kilo博客，免费获取新文章并支持我的创作。 旧有的交易原则 开源项目一直建立在一项简单的交易之上：... 
AI 在工程团队中的采纳深度：一个五阶段模型
Kilo 基于 300 万开发者、超 40 万亿 token 的遥测数据，总结出工程团队 AI 工具采纳的五阶段模型：从内联补全起步，到单文件编辑，再到需要代码库索引和 Plan Mode 支撑的多文件协作，再到并发 Agent 和持久会话。很多团队会长期停留在第二阶段，跨入第三阶段的关键不是热情而是上下文基础设施。 
Kimi K3 + Grok 4.5 构建的数据库与 Claude Opus 5 几乎相同，成本仅为 1/25
Kilo 实测用 Kimi K3 做规划、Grok 4.5 做实现，构建了一个嵌入式 KV 数据库，结果通过 64/65 项一致性检查，得分 93/100，而 Claude Opus 5 单独完成得 93/100，得分 98/100。组合成本仅 $1.27，Claude Opus 5 单独完成成本 $31.71，约为 4%。测试表明非旗舰开源模型组合在工程实现上已接近前沿模型水平。 
在不影响主设置的情况下运行第二个独立的 Kilo Code 实例
Kilo 的 CLI、VS Code 插件和 JetBrains 插件共享同一套全局配置与凭据，无法通过切换安装目录实现隔离。要运行第二个独立的 Kilo 实例，必须同时设置 XDG_CONFIG_HOME 和 XDG_DATA_HOME 两个环境变量，分别指向独立目录，否则凭据仍会共享。 Kilo 将配置（~/.config/kilo/）与凭据（~/.local/share/kilo/auth.json）分开放置，这是实现隔离的关键。仅修改 config 目录不会隔离 auth.json，因此必须双目录迁移。可通过 alias 封装命令简化操作，也可使用项目级.kilo/kilo.jsonc 实现仓库级区分，而非账号切换。 
适合任何硬件配置的最佳本地编码模型
Kilo Blog 在 2026 年 7 月评测了 9 款本地编码模型，按显存容量给出具体推荐：24GB 显存选 Qwen 3.6 27B，12-16GB 选 Gemma 4 12B Q4，48GB 或 64GB Mac 可跑 Qwen3-Coder-Next。所有模型均可通过 Kilo Code 和 Atomic Chat 离线运行，均为开源权重。 
组织领导者：你的 AI 政策能吸引顶尖人才吗？
Kilo Blog 指出：AI 使用政策直接影响顶尖人才选择，硬性限制 token 配额会引发人才流失。Gartner 预测 2028 年 AI 编码 token 成本将超过开发者平均薪资，6% 的组织每月人均支出超 2,000 美元。Uber 和 Microsoft 曾因预算超支实施单供应商上限，反而制造了新的工程与管理问题。 
远程会话中使用自定义模型
Kilo CLI 7.4.2 起远程会话支持调用本地模型，用户可在手机端选择 Ollama、LM Studio 或自托管端点模型。模型调用在本地机器完成，凭证不上传云端，Cloud session 仍默认使用 Kilo Gateway 模型但支持 BYOK 密钥路由至自有 provider。 
Two AI Stories, One Enterprise Trust Question
Two stories broke in AI this week that look unrelated. One is about a frontier model going off script during an internal safety test. The other is about a Chinese lab handing over the full weights to ... 
Auto Model vs Picking Your Own: We Tested Kilo Code's Router on a Backend Build
The pace of new model releases has made “which model should I use for this?” a real problem. Just in the last month we saw Claude Sonnet 5, Claude Fable 5, GPT-5.6, Grok 4.5, Gemini 3.6 Flash, and Kim... 
Build Your Castle, Dig Your Moat
“AI sovereignty” gets thrown around in enterprise slide decks until it stops meaning anything. A checkbox next to “on-prem.” As it’s been used historically, it’s a procurement term, a word borrowed to...
OpenRouter、Opus 5 与模型自由时代
Anthropic 发布 Claude Opus 5，价格减半且速度提升，推动高端推理模型成本下降。Stripe 传闻拟以 100 亿美元收购 OpenRouter，显示行业对模型路由和选择权的重视。Kilo 平台认为开发者生产力随模型选择增加而提升，封闭生态正在瓦解。 
微软 MAI 路由策略：模型选择 settle 的信号
微软在 Copilot 等产品中按任务路由流量至自研小模型，以降低成本并保留前沿模型处理复杂场景。 微软 Excel 团队报告其小模型在常见任务上与 GPT-5.6 持平，且无需最新加速器即可运行。 该策略核心在于建立基于实际工作负载的评估体系，确保质量底线，从而摆脱对单一供应商技术路线的依赖。 这种“按任务选模型”的路由机制比单纯追求模型参数规模更具工程实用价值。 