Skip to content

AI 资讯日报:2026-07-24 ​

岗位趋势 ​

FDE demand 2026 预计 +2,100%:仅 ~2,000 人具备 composite skill ​

FDE demand 2026 预计 +2,100%:仅 ~2,000 人具备 composite skill

TechCrunch 7-30 引用 Christian & Timbers 研究(7-24 行业讨论持续):美国仅约 **2,000** 工程师具备 sector know-how + gravitas + applied AI 的 FDE composite;demand 预计 2026 末 **+2,100%**。Q1 仅 5–10% 企业计划招 FDE pilot,Q2 跳至 **70%**——consulting firm 需 10× headcount(20–100 人 team)。Insurance/fintech/healthcare/gaming 均在建 in-house FDE team 防 knowledge 外流至 OpenAI/Anthropic。对 full-stack 想转 FDE:customer discovery + backend composite + on-site deployment 比 pure React depth 更稀缺。

来源: TechCrunch

行业动态 ​

DeepSeek V4 legacy alias HARD CUTOFF:15:59 UTC 生效 ​

DeepSeek V4 legacy alias HARD CUTOFF:15:59 UTC 生效

DeepSeek 官方 announcement **7-24 15:59 UTC**:`deepseek-chat` 与 `deepseek-reasoner` **fully retired**,请求返回 error 非 redirect。自 4-24 preview 起 alias 已 silently 路由至 V4-Flash——意味着 7-24 前 behavior 早已变,但 7-24 是 **code/config hard-break day**。Replacement:`deepseek-v4-flash`($0.14/$0.28 MTok,1M context)或 `deepseek-v4-pro`($0.435/$0.87);thinking 改为 request param。Developers Digest 7-25 确认:未迁移 team 的 **every API call failing**。对 monorepo:CI grep legacy name 应 fail build。

来源: DeepSeek API Docs

DeepSeek V4 GA:peak/off-peak 定价 announced,flat rate 仍有效 ​

DeepSeek V4 GA:peak/off-peak 定价 announced,flat rate 仍有效

NG Tech / Zenn 7-20–7-24 汇总:V4 从 preview 转 **GA**,同步 announce **Beijing peak window**(09:00–12:00、14:00–18:00 UTC+8)rate **double** off-peak baseline——但 **7-24 当日 official rate card 仍为 flat tier**,surcharge 无 confirmed start date。V3 off-peak discount 随 alias 一并退役。对 batch workload:宜 instrument request timestamp 为 future surcharge 做准备,但 **7-24 不必立即改 scheduler**。Migration 优先级:alias retirement > surge pricing speculation。

来源: NG Tech

热门项目 ​

MoonshotAI/Kimi-K3 ​

MoonshotAI/Kimi-K3

Moonshot GitHub repo **7-24 前** README/LICENSE 已就位,weights **目标 7-27**——2.8T MoE、1M context、Kimi K3 License(commercial use 允许,MaaS >$20M revenue 须 separate agreement)。Hugging Face `moonshotai/Kimi-K3` mirror 同步。对 front-end 全栈:Arena WebDev #1 relevant 但 self-host 需 ~64× H100——多数 team 7-24 仍走 API;repo star/watch 为 weights drop 做 CI template 准备(vLLM/SGLang config)。

来源: MoonshotAI/Kimi-K3

产品更新 ​

Claude Code v2.1.219:Opus 5 default + strictAllowlist sandbox ​

Anthropic release **7-24**:**Claude Opus 5** 设为 default Opus model(**1M context**);Opus 4.7 移出 fast mode。新 setting **`sandbox.network.strictAllowlist`**——仅 explicit listed destination 可 outbound,其余全部 block,防 compromised code / prompt injection exfiltration。新 **DirectoryAdded hook**;subagent 可 nested spawn 至 **depth 3**(原 default 1);`--forward-subagent-text` 在 stream-json 显示 nested agent output。Fix:text output loss、self-hosted runner restart 后 permission loss、model picker、screen reader。对 enterprise agent:strictAllowlist 是 production deploy 新 default 候选。

来源: 24 AI

DeepSeek migration one-liner:base_url 不变,仅 swap model name ​

DeepSeek migration one-liner:base_url 不变,仅 swap model name

Developers Digest checklist **7-24 cutoff day**:官方 advice 一行——"Keep base_url, just update model"。`https://api.deepseek.com` + 同 API key + OpenAI ChatCompletions format 不变。`deepseek-reasoner` traffic 额外加 `thinking: {"type": "enabled"}` + strip `reasoning_content` on replay。V4-Flash concurrency **2,500** vs Pro **500**——high-throughput pipeline 选 Flash 即使 quality gate 过 Pro eval。FIM completion 仅 non-thinking mode。对 emergency hotfix:单 env var `DEEPSEEK_MODEL=deepseek-v4-flash` 可 unblock 全队。

来源: Developers Digest

模型发布 ​

DeepSeek-V4-Flash vs Pro:AA Index 47 vs 52,cost 差 ~3× ​

DeepSeek-V4-Flash vs Pro:AA Index 47 vs 52,cost 差 ~3×

Developers Digest / Artificial Analysis **7-24 GA day** benchmark 参照:V4-Flash AA Intelligence Index **47**,full AA Index run cost **$113**;V4-Pro **52**,run cost **$1,071**。GDPval-AA agentic:Flash 1,388 vs Pro 1,554。Open weights MIT:Flash 160GB download vs Pro 865GB。Legacy `deepseek-chat` 用户 literal successor 是 **Flash non-thinking**——若需 Pro 是 deliberate upgrade,非 1:1 migration。对 cost-sensitive front-end agent:Flash + thinking enabled 先测,fail quality gate 再升 Pro。

来源: Developers Digest

Kimi K3 open weights 倒计时 3 天 ​

Moonshot 7-24 状态:**7-27** open weights target(3 天后)。K3 API 自 7-16 可用;Artificial Analysis Index 57,per-task ~$0.94。Hallucination rate ~51%(vs K2 ~31%)——production 须 monitoring。与 DeepSeek V4 GA 同日:国内 frontier 现 **API 双极**(DeepSeek 成本地板 + Kimi WebDev leaderboard)。对全栈 toolchain:7-24 是 DeepSeek migration deadline,K3 weights 是 secondary calendar——勿混为一谈。

来源: Moonshot Kimi Blog

论文研究 ​

Kimi K3 paper 预发布信号:KDA + Stable LatMoE scaling ~2.5× ​

Moonshot tech blog / 7-27 arXiv 预稿信息(7-24 社区引用):K3 基于 **Kimi Delta Attention + Attention Residuals** 改善 long-context information flow;**Stable LatMoE** 896 routed experts 激活 16/token,overall scaling efficiency vs K2 **~2.5×**。Post-training 含 general/agentic/coding 多 domain RL + multi reasoning-effort levels。对 ML-adjacent 全栈:paper 落地意味着 open weights 附带 inference recipe——关注 MXFP4 weights / MXFP8 activations quantization-aware training 细节以评估 local deploy feasibility。

来源: arXiv:2607.24653