On X
On — Nov 5
Open-weight model releases and coding-agent news lead the day. Cat Wu and Austin Lau announce moves to OpenAI; that story draws more attention on X than several brief items.
Kimi K3.1 Pro weights on Hugging Face#
Moonshot releases Kimi K3.1 Pro weights (2.8T total / 104B active / 1M context). Hosting and gateway day-0 support follow immediately.
Why it spiked. A large open model with usable specs is available to download and run.
- @Kimi_Moonshot Official weight release: “Releasing the model weights of Kimi K3.1 Pro… We'll keep advancing open intelligence.”
- @vercel_dev Day-0 AI Gateway support with Zero Data Retention for
moonshotai/kimi-k3.1-pro(and highspeed). - @Mr_Salio Specs: 2.8T total, 104B active, 1M context, now on Hugging Face.
Cat Wu (and Austin Lau) → OpenAI X-led#
Claude Code's product lead joins OpenAI Core Products. Anthropic's early growth hire posts the same night.
Why it spiked. Two known Anthropic people announce OpenAI moves on the same night.
- @_catwu “Some news: I'm joining OpenAI to lead product management for Core Products!”
- @helloitsaustin Confirms the OpenAI move; “Ant has better coffee/drinks and OAI has better snacks.”
- @DanielZambrini Notes OpenAI picking up Claude Code product and growth ahead of Anthropic's IPO window.
Conductor joining Meta (Cloud sunset)#
Conductor's founder says the multi-agent coding app is joining Meta's Superintelligence Labs. Cloud goes read-only Nov 13 and shuts down Nov 20; local builds continue.
Why it spiked. Active Cloud users have about two weeks before the service ends.
- @charlieholtz “I am very excited to share that Conductor is joining Meta!” (MSL with Nat Friedman and Alexandr Wang).
- @DonMark_EN Cloud read-only Nov 13, shut down Nov 20; local build planned to continue.
- @jondot If the harness lives in someone else's cloud, you get two weeks.
Claude Code — Haiku status updates (~99% cheaper)#
Claude Code background-subagent status updates no longer re-read the full chat every 30s. Haiku 5.5 writes the status line from short notes (~99% cheaper; update to v2.1.316+).
Why it spiked. Status ticks stop burning a large share of session quota.
- @ClaudeDevs Status updates “cost about 99% less” after the Haiku-from-notes change.
- @ClaudeDevs Sessions with background subagents use ~5% less limit; update to v2.1.316+.
- @nikolayxyz Old pattern resent the full chat every 30s just for a status line.
Tencent Hy4 — Artificial Analysis scores#
Artificial Analysis scores Hy4 at 35 on Index v5 (#2 China-based lab behind Kimi; ties Grok 4.7 xhigh).
Why it spiked. An independent score lands for a model people were already debating.
- @ArtificialAnlys Hy4 scores 35; #2 China-based lab; ties Grok 4.7 (xhigh).
- @superalesha China ladder: Kimi 46 → Hy4 35 → Qwen / DeepSeek / GLM; notes FrontierCode strength and caveats.
- @ItsmeAjayKV Places Hy4 #2 open-ish on the index behind Kimi, ahead of DeepSeek Flash and GLM.
Ripple-65B Preview (Ricursive)#
Ricursive releases Ripple-65B Preview (dense 65B, Apache-2.0) on Hugging Face and OpenRouter. Vendor claims TerminalBench 2.1 at 80.52 — treat as labeled until reproduced.
Why it spiked. New open coding weights ship with a high vendor benchmark claim.
- @BroadwakeAI Dense 65B, ~$32M R&D claim, TerminalBench 2.1 at 80.52; runs on DGX Spark / Mac-class setups.
- @bebuu_ai
ricursiveai/ripple-65b-previewlive on OpenRouter with 1M context. - @SISpotlight Frames it as a $32M open-weight coding model from a chip-design AI lab.
Cognition SWE-2 Flash#
Cognition ships SWE-2 Flash on Thinking Machines' Inkling-Small: about 5x cheaper and 3x faster than SWE-2. Weights stay gated, citing cyber risk.
Why it spiked. A faster, cheaper Devin model lands, but weights are not public.
- @cognition Introduces SWE-2 Flash; beats US open-weight peers on leading evals; partner program for weight sharing.
- @imjaredz “5x cheaper and 3x faster than the much larger SWE-2,” based on Inkling-Small.
- @superalesha Likes the model; disagrees with closing weights over the ExploitBench / cyber framing.