Skip to content

← AI Daily Picks

AI Daily Picks

Date7 November 2026 · SaturdayWindow2026-11-05 00:00 to 2026-11-07 08:00CoverageComplete

High-signal builder scan: Mira Murati’s Thinking Machines ships Inkling 2 Small (594B/17B MoE, AA Intelligence Index 31) — live on Tinker for inference/fine-tuning, with Apache-2.0 weights intended “next week” after safety tests; Ant Ling Ling-3.2-flash goes native multimodal (560B/25B, 1M context) with a two-week free API window (no published post-promo rate card for 3.2 yet); Baidu open-weights ERNIE-5.5-Flash-Base (280B/14B raw pretrain checkpoint, HF FP8+BF16); Moonshot formally launches Kimi Hosted Agents (Harness as a Service, sandbox ~¥0.072/hour); OpenAI Images 2.6 adds native GIFs + API Flare/Sunburst (~50% faster generation); Claude Code product lead Cat Wu announces she is joining OpenAI Core Products. Side note: Oumi FAiE productizes “the agent that builds agents.” Nov 6’s GPT-6 Cyber / Omni 1.2 / Claude Code −99% / Time per Task / Revolut GLM / Ripple-65B / Europe Blackwell talks are not reopened; Kimi K3.1 weights themselves are not restated.

1. US open-weight ladder, gen-2

Thinking Machines Inkling 2 Small: AA Index 31, Tinker first

  • What happened: Verified @thinkymachines on Nov 5 introduced Inkling 2 Small → https://thinkingmachines.ai/news/inkling-2-small/ . 594B total / 17B active MoE with image input and variable effort; Artificial Analysis Intelligence Index 31 (prior Inkling Small: 13). Available now on Tinker for playground / sampling / fine-tuning; weights stay closed at launch, with full Apache-2.0 release intended “next week” after external safety testing and red-teaming. Vendor tables cover agentic coding / vision / knowledge-work vs open Flash peers; alignment metrics improve vs prior Inkling (still trail some closed models).
  • Why it matters: US open-weight story adds a near-frontier MoE rung — and an explicit “host first, open after safety” path; teams that customize on Tinker can start today.
  • Source & engagement: thinkymachines — ~1.34M impressions, 4,990 likes, 480 reposts, 417 quotes, 238 replies, 2,677 bookmarks; follow-ups long-horizon, thanks. Secondary: Techmeme — ~5.4k.
  • Verification: verified — official blog + official X; board scores are vendor/AA snapshots; weights are not downloadable yet — don’t write “already open-sourced.”
  • Take: “Live on Tinker / Apache intended next week” is fine; “weights public” or “beats all closed frontiers” is not.

2. CN multimodal API free window

Ant Ling Ling-3.2-flash: 560B native image/video, free for two weeks

  • What happened: @AntLingAGI on Nov 5 launched Ling-3.2-flash — 560B total / 25B active, 1M context, native text/image/video; focus areas: complex visual understanding, long-horizon multimodal agents, 3D, medical. Vendor release benches vs Opus 5 (max): OmniDocBench1.5 92.21 vs 90.56, ZEROBench(Sub) 43.71 vs 39.52, Vision2Web 85.86 vs 80.20. Docs: https://developer.ant-ling.com/ ; changelog states first 2 weeks free on the API. The official price table has no 3.2 row yet — circulating $0.30/$0.90 figures are Ling-3.1-flash OpenRouter list rates; do not budget 3.2 with them.
  • Why it matters: A cheap try-window for vision-in-the-agent-loop; post-promo unit price unknown — explore capability, don’t lock a long-term cost model.
  • Source & engagement: AntLingAGI — ~356k impressions, 927 likes, 97 reposts, 111 quotes, 72 replies, 217 bookmarks.
  • Verification: verified — official X + developer docs; three benches are vendor self-report; list price pending.
  • Take: “Two-week free / no 3.2 rate card yet”; don’t paste 3.1’s $0.30/$0.90 as 3.2 pricing.

3. Raw base weights

Baidu ERNIE-5.5-Flash-Base: 280B/14B pretrain-endpoint checkpoint

  • What happened: Verified @Baidu_Inc on Nov 5 released ERNIE-5.5-Flash-Base — sparse MoE 280B total / 14B active, pretrained from scratch on ~33.5T tokens; train context up to 128K, fine-tune/infer up to 1M. Weights are the end-of-pretrain checkpoint before mid/post-training, FP8 + BF16 on HF: https://huggingface.co/baidu/ERNIE-5.5-Flash-Base ; product: https://ernie.baidu.com/release/ernie-5-5-flash-base/ . License is Baidu’s custom agreement (not Apache): excludes EU/UK/South Korea; related-product annual revenue above ~$50M needs prior written authorization — follow LICENSE.
  • Why it matters: Rare “raw base meat” from a major lab — a starting point for SFT/RL and domain adaptation; geo/commercial license gates must appear in any “open” headline.
  • Source & engagement: Baidu_Inc — ~238k impressions, 335 likes, 47 reposts, 33 quotes, 48 replies, 76 bookmarks; CN secondary NFT_Chen — ~25k / 77 likes.
  • Verification: verified — official X + site + HF card/LICENSE; board scores are vendor self-report.
  • Take: “Open weights / custom license / pure Base” — not “fully Apache open source” or “freely commercial in the EU.”

4. Harness becomes a cloud bill

Kimi Hosted Agents: managed runtime at ¥0.072/sandbox-hour

  • What happened: Moonshot formally launched Kimi Hosted Agents on Nov 5 — packaging the agent Harness (sessions, tools, sandboxes, failure recovery) as a cloud service (Harness as a Service) for 24/7 long-horizon and multi-agent work; docs: https://platform.kimi.com/docs/hosted-agents/overview ; news: https://www.kimi.com/news/kimi-hosted-agents . Includes 140+ Skills, 30+ plugins and finance-oriented templates; optional Inference Hook inspects inputs/tool returns before each model call. Pricing: sandbox ¥0.072/hour + model tokens + plugin calls (see pricing page).
  • Why it matters: After same-week weight drops, the narrative shifts to “who runs your agent plumbing” — cost/ops contrast vs DIY or open harnesses.
  • Source & engagement: Official EN X was thin this window; CN secondaries xiax0603 — ~1.1k; leo114119 — ~1.8k; prefer product docs.
  • Verification: verified — official product/pricing pages; X mostly secondary digests.
  • Take: “Managed harness + sandbox rate” — don’t read “no ops” as “no model/plugin spend.”

5. Image product adds motion

OpenAI Images 2.6: native GIFs + Flare/Sunburst API

  • What happened: Verified @OpenAI late Nov 4 (Asia Nov 5) launched ChatGPT Images 2.6 — native GIFs, reusable reference images; API models GPT-Image-2.6 Flare (as little as ~6s) and Sunburst (higher quality / edit consistency), with ~50% faster image generation per OpenAI. Product: https://openai.com/index/introducing-images-2-6/ . Available across ChatGPT/Codex tiers on web, desktop, mobile.
  • Why it matters: Short motion no longer needs a second external tool; pairs with yesterday’s Omni 1.2 video pricing as creative-media cost context — not covered in the Nov 6 daily.
  • Source & engagement: OpenAI main — ~3.31M impressions, 21,759 likes, 1,875 reposts, 1,915 quotes, 1,174 replies, 4,422 bookmarks; GIF — ~360k; API — ~168k.
  • Verification: verified — official blog + official X.
  • Take: “Native GIF / two API SKUs” — don’t conflate with video-length caps.

6. Product talent move

Cat Wu: Claude Code PM joins OpenAI Core Products

  • What happened: @_catwu on Nov 5 — left Anthropic last Friday; joining OpenAI to lead product management for Core Products, working with @thsottiaux and others on user-controlled agents. She previously product-led Claude Code / Cowork. Same-window community lists also cite Austin Lau and others leaving/joining (secondary tallies, not an Anthropic org chart).
  • Why it matters: People signal for the coding-agent product war — roadmap tea, not a model drop.
  • Source & engagement: _catwu — ~4.01M impressions, 17,306 likes, 742 reposts, 1,427 quotes, 1,995 replies, 1,944 bookmarks; secondary pradeepXkapoor — ~18k.
  • Verification: verified (first-person announcement) — scope follows her post; “exodus” lists are community compilations.
  • Take: “Self-announced move” — don’t write “Claude Code is dead” or invent motives.

7. Side: the agent that builds agents

Oumi FAiE: eval → data → train → deploy loop

  • What happened: Verified @oumi_ai and co-founder @Koukoumidis on Nov 4 launched FAiE (Frontier AI Engineer) — describe a task in plain language; it designs evals, diagnoses failures, curates data, picks training (SFT/DPO; GRPO private preview) and runs experiments; supports Bring Your Own Harness. Platform: https://platform.oumi.ai/ ; event https://oumi.ai/events/fall26 . Marketing claim: ~15 minutes of user input + a few hours of automated engineering → specialized weights.
  • Why it matters: Productizes AI engineering itself — good newsletter footnote, not a proven production SLA.
  • Source & engagement: oumi_ai — ~129k; Koukoumidis — ~1.19M / 309 likes.
  • Verification: verified — official accounts/stream; “15 minutes” is vendor claim pending independent case studies.
  • Take: “Workflow product + BYO harness” — don’t write “replaces every ML engineer.”