Skip to content

AI 资讯日报:2026-07-28 ​

岗位趋势 ​

FDE posting +729% YoY:Tredence 称 enterprise AI 最后一公里 ​

Tredence blog 7-27(7-28 引用):Forward Deployed Engineer posting 2025-04 ~643 → 2026-04 ~5,330(+729% YoY);2026-05 39 家 AI 公司共 224 个 open FDE role。Median base $127K–$173K,frontier lab total comp mid-level ~$385K、principal 可达 seven figures。FDE 非 staff augmentation——驻客户侧 product engineering,handoff chain 从 business → consultant → data → eng → deploy 压缩为单一 embedded team。对全栈 lateral move:guardrails + RAG + eval + customer discovery 组合技能。

来源: Tredence

FDE 是 AI-native SDLC 2.0 角色:review agent + human gate 并行 ​

Tredence 7-27 补充:其 FDE 模型强调 AI-Ninja coding fluency——discovery/design/coding/testing/semantic modeling/documentation/ops 全链路在 AI-native environment 完成,非传统 SDLC 上 bolt-on Copilot。Harness 含 review agents、pre-commit checks、human gates;accelerators 覆盖 40–60% scaffolding,engagement 间 compound。Platform-agnostic(Databricks/GCP/Azure/Snowflake/AWS + leading model providers)。对 frontend/fullstack:FDE 不是纯 backend infra 岗,semantic layer + agent guardrail 设计是 core deliverable。

来源: Tredence

行业动态 ​

Kimi K3 active params 是 104B 非 ~50B:devoriales 纠正社区误算 ​

Kimi K3 active params 是 104B 非 ~50B:devoriales 纠正社区误算

devoriales 7-27/28:Hugging Face community post 用 16/896 expert ratio 推算 ~50B active——**错误**,遗漏 2 shared experts 与每 token 都运行的 dense attention stack(69 KDA + 24 Gated MLA layers)。Moonshot model card 明确 **104B activated per token**。影响:infra sizing、latency 估算、与 GPT/Claude active param 对比均须用 104B baseline。AA Intelligence Index 57、hallucination rate 51%(K2 ~39%)仍须在自有 eval 验证。

来源: devoriales

热门项目 ​

MoonshotAI/Kimi-K3 inference recipes ​

MoonshotAI/Kimi-K3 inference recipes

Weights 发布次日 community 聚焦 deployment:repo 推荐 vLLM、SGLang、TokenSpeed;896 expert routing + KDA state + MLA KV cache 联合调度是 day-1 痛点。MXFP4 expert weights / MXFP8 activations 使 ~1.4TB resident footprint 可 datacenter serve,但非 desktop 场景。MLOps/FDE 团队可先 fork inference config template 再接入 enterprise gateway——OpenRouter `moonshotai/kimi-k3` 作 fallback 对照。

来源: MoonshotAI/Kimi-K3

产品更新 ​

Kimi K3 day-0 上 Fireworks / Together / OpenRouter ​

Kimi K3 day-0 上 Fireworks / Together / OpenRouter

Moonshot launch-day partners + OpenRouter 7-28 路由矩阵:Fireworks `accounts/fireworks/models/kimi-k3`(serverless + dedicated + fine-tuning,US-hosted zero retention option);Together serverless/dedicated;OpenRouter `moonshotai/kimi-k3` 聚合 billing。统一定价 $3 input / $15 output / $0.30 cache read per MTok。AWS Bedrock、Azure Foundry、Google Vertex 尚未集成(Memeburn 7-28 引用 Bloomberg Intelligence)。对全栈:Vercel AI SDK / OpenAI-compatible client 改 model ID 即可 day-0 接入,无需等 hyperscaler。

来源: OpenRouter

Kimi K3 provider 矩阵:Modal / Baseten / DigitalOcean 同日上线 ​

Kimi K3 provider 矩阵:Modal / Baseten / DigitalOcean 同日上线

Developers Digest 7-27 snapshot(7-28 核验):除 Fireworks/Together 外,Modal(73 tps)、Baseten、DigitalOcean Serverless、SiliconFlow、Cloudflare Workers AI、RunPod 等均 listing K3。OpenRouter 当前仅 1 upstream provider——consolidated key 方便但无 meaningful failover。Nebius eu-north1 FP4 部署 signal:同价 3–4× faster(Teortaxes 观察)。Frontend team 选 provider 时须 weigh latency SLA vs data residency vs cache hit rate(Moonshot upstream 91% cache share on OpenRouter)。

来源: Developers Digest

模型发布 ​

Kimi K3 open-weights 定位:AA Index 57,WebDev Arena #1 但 hallucination 回归 ​

Kimi K3 open-weights 定位:AA Index 57,WebDev Arena #1 但 hallucination 回归

devoriales / Artificial Analysis 7-28 汇总:K3 Intelligence Index 57,overall #4(次于 Fable 5、GPT-5.6 Sol variants),open-weights #1/#3 取决于榜单 cutoff。Arena.ai blind Code WebDev #1 relevant for frontend。但 hallucination rate API 测量 ~51% vs K2.6 ~39%——weights 落地前 hosted API eval 与 self-host eval 可能 diverge。Production 须跑自有 regression + 监控,不可仅信 marketing benchmark。

来源: devoriales

论文研究 ​

Kimi K3 Stable LatentMoE + KDA:2.5× scaling efficiency over K2 ​

arXiv 2607.24653 tech report 7-27:KDA 改善 long-context information flow;Attention Residuals 改善 depth-wise flow;Stable LatentMoE 有效激活 896 routed experts 中 16/token。Algorithm-system co-design 覆盖 expert-parallel training、million-token agentic RL with persistent rollout/sandbox states。Post-training RL 跨 general/agentic/coding + 多 reasoning-effort。Research/engineering 交界:paper infra section 对 planning self-host 的 FDE 团队是直接 sizing 输入。

来源: arXiv:2607.24653