Skip to content

← AI Daily Picks

AI Daily Picks

Date5 November 2026 · ThursdayWindow2026-11-03 16:00 to 2026-11-05 08:20CoverageComplete

High-signal builder scan: Moonshot drops full Kimi K3.1 Pro weights on the promised day (HF moonshotai/Kimi-K3.1-Pro — 2.8T MoE / 104B active / 1M context); Tencent Hy4 GA with first-party vision + Computer Use — API live on Tencent Cloud / OpenRouter, weights “in the coming days”; Anthropic open-sources FlashProve (zk proofs that a reply came from committed weights); OpenAI’s eng blog details monorepo Git backend Pando (~5× / ~80% median push latency cut, ~1 PB Git data/day). Side note: Japanese eng YASAKANI self-reports ~1,220 tok/s single-stream on one RTX 5090 for Qwen3.8-27B. Nov 4’s weight-space programming / Hume audio / Pulse·Crouton are not reopened.

1. Open-weight calendar lands

Kimi K3.1 Pro: weights on Hugging Face the day they promised

  • What happened: Verified @Kimi_Moonshot posted Nov 4 16:00 UTC “Releasing the model weights of Kimi K3.1 Pro” → https://huggingface.co/moonshotai/Kimi-K3.1-Pro . Model card: MoE 2.8T total / 104B active, context 1,048,576, native multimodal (MoonViT-V2), MXFP4/MXFP8; recommended engines vLLM / SGLang / TokenSpeed. Databricks CEO Ali Ghodsi posted “Zero Day Launch”; @databricks claimed day-0 availability. License is the Kimi K3.1 Pro License (community read: large-scale API resale / high MAU carve-outs — check the HF license file; don’t call it plain MIT).
  • Why it matters: Yesterday’s brief still said “promised by Nov 4”; today it’s a downloadable repo — CN open-weight calendars moved from slogan to artifact. Local run cost stays brutal (~1.56TB shards / multi-GPU class per community), but evals / distillation / harness A-Bs can start.
  • Source & engagement: Kimi_Moonshot — ~363k impressions, 5,928 likes, 629 reposts, 335 quotes, 277 replies, 748 bookmarks; alighodsi — ~3.3k views, 78 likes.
  • Verification: verified — official X + live HF card; bench numbers are vendor self-report; license details follow the HF License file.
  • Take: “Weights shipped” is fine; “everyone can run it locally” is not. Don’t treat internal boards (Kimi Code Bench, etc.) as independent leaderboards.

2. CN flagship GA

Tencent Hunyuan Hy4: vision + computer use; weights “coming soon”

  • What happened: Verified @TencentHunyuan announced Hy4 GA on Nov 4 — “new with vision,” “Firmly among the leading open models”; product page https://hy.tencent.ai/research/hy4 ; Tencent corporate post says weights + system card in the coming days, with vLLM / SGLang ecosystem support. First Hy model with vision understanding + Computer Use; live on Tencent Cloud, OpenRouter, CodeBuddy, WorkBuddy, Hy AI Studio. Secondary community numbers (770B MoE, OSWorld splits, etc.) are mostly vendor/relayed — check the product page before quoting.
  • Why it matters: Same window as “Kimi weights live / Hy4 API first, weights later” — separate open-source rhetoric from downloadable facts. Computer Use + office/finance/game framing is the product story.
  • Source & engagement: TencentHunyuan — ~449k impressions, 2,652 likes, 303 reposts, 268 quotes, 231 replies, 518 bookmarks; TencentAI_News — ~39k / 887 likes; testingcatalog — ~9.6k / 123 likes; TencentGlobal.
  • Verification: verified — official accounts + Tencent.com post + product page; “leading open” is vendor claim; weights not observed in this window → write “coming soon.”
  • Take: Don’t write “Hy4 weights are downloadable.” WorkBuddy free windows / TokenHub discounts follow product pages. “Users prefer congested DeepSeek” chatter is anecdotal.

3. Provable inference

Anthropic FlashProve: cryptography that a reply came from committed weights

  • What happened: Verified @AnthropicAI introduced FlashProve — cryptographic techniques so developers can verify a response came from a specific committed model, substantially more efficient than prior open-source approaches. Research: https://www.anthropic.com/research/provable-inference ; code Apache-2.0: https://github.com/anthropics/FlashProve . Reported demo: Llama 4 Maverick (~400B MoE, 4-bit) on one H100 proves at ~110 tok/s, ~19 MB proof per 1,024-token request — about 70–95× the GPU time of running the same model on vLLM. Use cases: harder chat hijacks; safety evaluators checking the tested model matches production.
  • Why it matters: Closes part of the “did the router swap my model?” trust gap with a runnable open prototype — narrative shifts from vendor trust to auditable commitments.
  • Source & engagement: AnthropicAI — ~607k impressions, 4,998 likes, 519 reposts, 459 quotes, 603 replies, 2,057 bookmarks.
  • Verification: verified — official research page + GitHub; throughput/proof size are vendor self-report; labeled research prototype, not a prod SLA.
  • Take: Don’t write “Claude API now ships proofs by default”; hiding architecture isn’t a goal; 70–95× overhead means near-term fit is compliance/audit, not default chat.

4. Source control for the agent era

OpenAI Pando: a Git backend rebuilt for Codex-scale concurrent pushes

  • What happened: OpenAI eng blog (2026-11-02) Scaling Git at OpenAI: https://openai.com/index/scaling-git-at-openai/ — new Git backend Pando, keeps Git protocol/client, tuned for engineers + Codex agents pushing/fetching hard. Official: median push latency down ~5× (~80%), nearly 1 PB/day Git data; packfiles in blob storage, per-branch coordination, pack reuse across clients; ~10 weeks to prod with Codex help, serving since April. X digests kept circulating Nov 3–4.
  • Why it matters: Agent coding bottlenecks are moving from “the model” to “repo infra” — pairs with Anthropic’s Scion narrative (new backend vs new protocol).
  • Source & engagement: Prefer the OpenAI blog; X secondary e.g. AINativeF.
  • Verification: verified — official eng blog; X posts are secondary.
  • Take: Pando is internal; no public open-source pledge seen this pass. Don’t write “drop-in for your GitHub.” Stick to “~5× / ~80%” as in the post.

5. Inference engine side note

YASAKANI: self-reported 1,220 tok/s on one 5090 (Qwen3.8-27B)

  • What happened: Japanese developer @yasakani_dev posted custom engine YASAKANI, claiming ~1,220–1,260 tok/s single-stream decode for Qwen3.8-27B on one non-overclocked RTX 5090, and wins vs SGLang / vLLM / llama.cpp on their published single-card records (up to ~2.5–2.86×). EN secondary: Oluwaphilemon1; quote chain sample via crazyMLguy.
  • Why it matters: Consumer single-GPU tok/s races still shape local-agent latency narratives — needs independent repro.
  • Source & engagement: author post + secondaries (~3.7k on the EN sample).
  • Verification: author-claim — self-measured; no third-party blind bake-off seen this window.
  • Take: Write “self-reported”; don’t write “vLLM is obsolete.” Speculative-decoding / batch knobs, if unpublished, block fair cross-engine comparisons.

6. CN ecosystem (light)

License wording vs “open”; DeepSeek V5 × Ascend still planning coverage

  • What happened: Community reads of the Kimi license as “use/tune/sell OK, but large API resale / high MAU need extra deals” remain author-claim/community interpretation — HF license file wins. Separate reporting that DeepSeek may pretrain V5 mainly on Huawei Ascend 950 (~2027) lacks official confirmation from either party in this window. miHoYo’s multi-year AI spend headlines are finance/reporting, not a model drop.
  • Why it matters: On hot open-weight days, license + silicon supply often decide private-stack fit more than leaderboard rows.
  • Source & engagement: license chatter e.g. vreskai; Ascend notes via secondary digests.
  • Verification: license read author-claim; DeepSeek×Ascend unverified vs official; miHoYo reporting-grade.
  • Take: Don’t write planning as “training already underway.”