1
阶跃 Step 5 Preview 大模型权重开放，支持 100 万 token 上下文窗口
IT之家 10 月 15 日消息，阶跃今天正式释放 Step5Preview 权重。本次开放：BF16 与 FP8 模型权重、技术博客、配套开源开发工具。 据官方介绍，Step 5 Preview 面向长程 Agent 任务而设计，采用稀疏 MoE 架构，总参数量 600B、每 token 激活参数…
2
MAI Code 1.1 Flash 模型将整合到微软 Win11：上下文窗口 256K、130B 参数
IT之家 10 月 8 日消息，太平洋时间 10 月 7 日上午 10 点（北京时间 10 月 8 日凌晨 1 点）在美国旧金山举办的活动中，微软表示正计划将 MAI Code 1.1 Flash 模型（包含 1300 亿个参数）引入 Windows 11 设备。 作为 6 月发布的 MAI-Cod…
3
Show HN: Neither HTML nor Markdown is enough: a way out of the AI doc dilemma
AI文档格式为何不够用？从Claude Code的HTML经验出发，提出兼顾注意力与持久记录的新思路。
A recent post from Anthropic's Claude Code team ("The Unreasonable Effectiveness of HTML", https://thariqs.github.io/html-effectiveness/) argued that …
4
Show HN: Ollama served my 40k-context model at 4k, silently
别被Ollama悄悄缩短上下文坑了，这个小工具帮你检测真实窗口大小。
Article URL: https://github.com/Bigbonus/ollama-context-window-check Comments URL: https://news.ycombinator.com/item?id=49372235 Points: 1 # Comments:…
5
Token Optimization and Context Window Management in Multi-Agent AI Workflows
多智能体成本与延迟双降的实战框架，生产级仪表板赋能上下文优化，值得实践者深读。
arXiv:2608.17188v1 Announce Type: cross Abstract: Multi-agent AI workflows are limited not only by model quality but by token cost, latency, and conte…
6
Forget Total Recall. Give Your AI Agent Selective Memory with OpenSearch.
别让长文本淹没重点，用OpenSearch为AI智能体打造选择性记忆，精准调用所需信息。
Imagine handing an LLM the full text of Anna Karenina and asking what Levin thinks about farming. The model has the information somewhere in its windo…
7
AGENTS.md, explained for teams that actually ship
给真正交付的团队：AGENTS.md不是README，而是代码库地图。用5-10条要点、300行上限，让AI代理精准理解你的业务逻辑。
By the end of 2025, AI coding agents stopped being a toy. Claude Code, Codex, Cursor, Copilot and opencode are now doing real work in real repos — and…
8
What Is Context Rot? Why AI Agents Degrade Mid-Session
越聊越笨不怪你：上下文腐烂正悄悄拖垮AI Agent的会话质量，开发者的信任危机值得警惕。
Originally published at getunblocked.com on August 10, 2026. Context rot is the gradual degradation of an LLM's output quality as its context grows — …
9
Show HN: Ctxdiff – Git diff for your LLM agent's context window
像Git diff一样对比LLM agent的上下文窗口，精准定位每次回合中破坏缓存的“罪魁祸首”字符。
Article URL: https://github.com/salmanzafar949/ctxdiff Comments URL: https://news.ycombinator.com/item?id=49070029 Points: 1 # Comments: 2
10
Planning with Transformers: Chain of Computation and Structured Context Windows
Transformer新突破：链式计算与结构化上下文窗口，让AI规划能力再升级！
arXiv:2607.17710v1 Announce Type: new Abstract: Large Language Models (LLMs) have had a remarkable impact across many areas of machine learning. Howev…
11
OpenAI reduces Codex Model Context Size from 372k to 272k
OpenAI将Codex模型上下文压缩近27%，通过优化降低显存占用与推理成本，开发者值得关注这一效率改进。
Article URL: https://github.com/openai/codex/pull/33972/files Comments URL: https://news.ycombinator.com/item?id=48965850 Points: 281 # Comments: 134
12
报告称 Claude Code 处理用户请求前消耗 Tokens 量约为 OpenCode 的 4.7 倍
Claude Code处理请求前Token消耗高达OpenCode近5倍，上下文窗口被挤占，开发者面临成本与效率挑战。
IT之家 7 月 14 日消息，科技媒体 systima 于 7 月 12 日发布博文，对比测试指出在相同提示词、相同 AI 模型情况下， 在处理用户请求前，Claude Code 消耗的 Tokens 数量约为 OpenCode 的 4.7 倍。 该媒体对比测试 Claude Code 2.1.2…
13
Self-Guided Test-Time Training for Long-Context LLMs
自我引导的测试时训练，专治长上下文LLM的准确率退化难题。
arXiv:2607.09415v1 Announce Type: cross Abstract: Long-context processing has become increasingly important for large language models (LLMs), but simp…
14
I Did the Math on GPT-5.6. The $2.50 Terra Tier Is the One I'd Ship First.
GPT-5.6 Terra tier仅$2.50/百万输入token，1M上下文窗口直击生产负载痛点，性价比碾压Sol。
GPT-5.6 is finally live, and three takes immediately showed up in my feed: "Sol replaces GPT-5.5 everywhere." "The API still isn't broadly available."…
15
谷歌 Gemini 3.5 Pro 曝 200 万 Tokens 上下文，前端部分 AI 测试赶超 Claude Fable 5
谷歌Gemini 3.5 Pro曝光200万Tokens上下文与深度思考模式，前端测试性能已超越Claude Fable 5。
IT之家 7 月 7 日消息，消息源 @HarshithLucky3 昨日（7 月 6 日）在 X 平台发布推文，爆料称谷歌计划 7 月 17 日发布 Gemini 3.5 Pro 模型， 支持 200 万上下文窗口，引入全新“深度思考”推理模式等。 定位方面，消息称 Gemini 3.5 Pro …
16
openclaw 2026.7.1-beta.2
更新支持GPT-5.6与百万级上下文窗口，openclaw让AI工作流更强大
Highlights OpenAI GPT-5.6 support: OpenClaw now recognizes the GPT-5.6 model family across catalog, capability, and runtime selection paths. ( #98333 …
17
How AI Agents Manage Memory and Avoid Forgetfulness
一张图讲清AI代理如何用外部记忆避免“失忆”，直击上下文窗口的致命短板。
In this article, we will try to understand how that architecture gets built, from the constraint that forces it to exist all the way to the tradeoffs …
18
Show HN: Statuslin.es – a community library of custom Claude Code status lines
自定义Claude Code状态行，实时显示上下文窗口使用率与模型切换，开源社区共享
A little while ago I was searching for a nice status line to use in Claude Code and was surprised that no community library type website existed for p…
19
Don't trust large context windows
大模型上下文窗口并非越大越好，注意力在10万token后断崖式下降，编码代理人尤其危险。
Article URL: https://garrit.xyz/posts/2026-05-06-dont-trust-large-context-windows Comments URL: https://news.ycombinator.com/item?id=48524620 Points: …
20
AI Has Amnesia. Here's Every System Built to Fix It
AI失忆症不再是难题，这篇全方位盘点所有修复方案，一网打尽各大系统与技术。
Article URL: https://medium.com/@alanayalag/your-ai-has-amnesia-heres-every-system-built-to-fix-it-ad7dee117a75 Comments URL: https://news.ycombinator…