AI 资讯日报：2026-07-26
岗位趋势
AI labs 吸引力超 Big Tech：Anthropic/OpenAI interview prep demand 最高
Pragmatic Engineer 2026 job market data（7-26 引用）：**Anthropic 成 job candidate interview prep 需求最高雇主**，OpenAI 并列——top AI labs 吸引力 **超过 Big Tech**。AI engineering comp > software engineering comp（80th percentile $250K+ base + higher equity）。Big Tech AI engineering headcount 显著增长，software engineering opening 仅 +7% YoY。对 senior full-stack 考虑 offer：AI lab 的 equity upside + agent/product 曝光 vs Big Tech 的 stability——7-26 Kimi K3 + DeepSeek V4 双线 open weights 亦扩大 "lab-adjacent" employer 选择（Moonshot/DeepSeek hiring）。
行业动态
Kimi K3 open weights eve：2.8T MoE 明日 target，Together/Modal day-0 就绪
Moonshot platform + community tracking **7-26 eve**（official target **7-27**）：K3 2.8T MoE、1M context、104B activated params——**largest open-weight release tracked in 2026 frontier cycle**。API 自 7-16 运行；Artificial Analysis Index 57；Arena Code WebDev #1。Self-host realistic threshold ~64× H100（community estimate）——多数 front-end team 仍 API/hosted path。Together AI / Modal 据 signal 准备 day-0 inference endpoint。License：Kimi K3 License（commercial OK，MaaS >$20M revenue 须 agreement）。对全栈：**7-26 完成 API eval baseline**，weights 落地后同一 harness 对比 hosted vs self-hosted。
DeepSeek post-GA 48h：零 legacy alias 404 ticket 应已清零
deepseek.ai blog / Developers Digest **7-26 status**：alias cutoff 48h 后——production 仍发 `deepseek-chat` 的 integration **100% failing**。Surge pricing **announced 未 active**；flat rate unchanged。Next action item 从 emergency migration 转向 **Flash vs Pro explicit choice + cache optimization**。Anthropic-compatible endpoint（见产品更新）为 Claude SDK standardized team 提供第二 migration path。对 on-call：7-26 应 zero P1 "model not found"——若仍有，根因通常是 eval harness/notebook 漏 grep。
来源： DeepSeek.ai Blog
热门项目
MoonshotAI/Kimi-K3
GitHub/Hugging Face **7-26 eve**：`MoonshotAI/Kimi-K3` repo LICENSE + inference code 就位，weights **7-27 target**。Kimi K3 License 允许 modify/commercial deploy；100M+ MAU 或 $20M+ monthly revenue product 须 prominent branding。MXFP4 weights / MXFP8 activations quantization-aware training——local deploy 可行性取决于 GPU fleet。对 front-end 全栈：star repo + prepare vLLM/SGLang config template；WebDev Arena #1 relevant 但 production 须自有 hallucination monitoring（AA 报告 ~51% rate）。
产品更新
DeepSeek Anthropic-format endpoint：Claude SDK 两 env var 迁移
Developers Digest / DeepSeek official Anthropic API guide **7-26 重点**：第二 API surface **`https://api.deepseek.com/anthropic`**——Anthropic Messages format。迁移 Claude SDK standardized codebase：`ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic` + `ANTHROPIC_API_KEY=<DeepSeek key>`。Auto mapping：claude-opus → `deepseek-v4-pro`，claude-sonnet/haiku → `deepseek-v4-flash`，unmapped → Flash default。Support：streaming、temperature、stop、top_p、tool use、thinking mode（`budget_tokens` ignored）。**Not supported**：images、documents、code execution results、MCP content types。对 Claude Code 用户：最快 V4 path；production 宜 pin explicit model name 防 upstream mapping change。
DeepSeek Anthropic bridge：Claude Code 用户最低 friction V4 路径
DeepSeek official docs `guides/anthropic_api`（7-26 引用）：V4 preview 起 marketed Claude Code/OpenClaw/OpenCode 集成——Anthropic-compatible endpoint 使 **Claude SDK / Claude Code config 用户无需 rewrite client code**。Trade-off：multimodal + MCP content 不可用——pure text coding agent 场景 fit，document-upload RAG agent 不适用。与 7-24 Claude Code Opus 5 default + strictAllowlist 对比：DeepSeek path 是 **cost floor alternative**，非 feature parity。Team 可 A/B：同 repo 跑 Claude Code (Opus 5) vs DeepSeek-bridge (V4-Pro) eval。
模型发布
Kimi K3 open weights 倒计时 1 天
Moonshot **7-26 eve**：official open weights target **7-27**（1 天后）——2.8T total / 104B activated MoE，native vision，1M context。Paper arXiv:2607.24653 将同步（submitted 7-27）。Benchmark claim：frontier-level long-horizon coding/agentic/knowledge/reasoning/vision；overall 仍 trail Fable 5 / GPT-5.6 Sol 但 **lead other open + proprietary models in evaluated suite**（paper abstract）。OpenRouter `moonshotai/kimi-k3` API 已 running 10 days——weights 落地是 pricing/power 分水岭（super-cheap Chinese AI era 可能结束）。对全栈：7-26 最后一轮 API baseline capture。
来源： arXiv:2607.24653
DeepSeek V4-Flash：legacy deepseek-chat 的 literal 1:1 successor
DeepSeek official + Developers Digest **7-26**：post-cutoff canonical mapping——`deepseek-chat` → **`deepseek-v4-flash` non-thinking**（behavior 4-24 起已 silently 切换，7-24 起 name hard-break）。$0.14/$0.28 MTok，1M context，tool calls + thinking param unified。Anthropic endpoint 上 claude-sonnet/haiku class → Flash。对 monorepo standardized on `deepseek-chat` 的 **fallback chain**：update to explicit `deepseek-v4-flash` 即可 restore——无需 base_url/key rotation。Surge pricing watch continues but flat rate holds 7-26。
论文研究
arXiv:2607.24653 Kimi K3：2.8T MoE + million-token agentic RL infrastructure
Moonshot **arXiv 2607.24653**（7-27 submit，7-26 预阅 circulating）：Kimi K3 2.8T MoE / 104B activated，KDA + Attention Residuals + Stable LatMoE（896 experts, 16 active/token）。Infrastructure：algorithm-system co-design for KDA、balanced expert-parallel training、**million-token agentic RL with persistent rollout and sandbox states**。Release full weights for research + deployment acceleration。对 agent infra 工程师：paper 的 sandbox-state persistent RL 设计 relevant 于 long-horizon coding agent eval——K3 DeepSWE 67.3% 须 cross-check public leaderboard harness vs KimiCode self-report。
来源： arXiv:2607.24653