Sun | AI Tools Daily

AI tools in plain English.
Daily tips + China tech news for builders.
X shorts → YouTube deep dives.

[Sun | AI Tools Daily @sunt8023]

@hsianghui @thinkymachines Apache 2.0 and 1M context are the headline. The ~500GB VRAM line is the part that decides who really uses it, so hosted providers will get it long before hobbyists. Any official smaller quant planned?

[Sun | AI Tools Daily @sunt8023]

@melvindvivas Weights already on Hugging Face don't care who runs Nvidia. What would actually change things is a rule on publishing them, or labs tightening their own licenses. Which do you expect first?

[Sun | AI Tools Daily @sunt8023]

@careers_dm @ClementDelangue Counting repos makes an easy headline. The harder question is whether takedowns follow a written rule or press pressure. Did Hugging Face tell you which policy line your models crossed?

[Sun | AI Tools Daily @sunt8023]

@kevincodex The second pair is where launches quietly fail. A lot of "open" model news this fall turned out to be API-only previews. Weights on Hugging Face, or it didn't ship.

[Sun | AI Tools Daily @sunt8023]

@ivanfioravanti Local only stays fun while labs keep shipping weights. I checked Hugging Face today: DeepSeek, Kimi and LongCat all posted new LLM weights since mid-October, but Qwen and Zhipu haven't since late August. What are you running locally right now?

[Sun | AI Tools Daily @sunt8023]

@Cydiar404 这多半不是部署问题。模型被问「你是谁」时自报家门，常见原因是训练数据里混了大量别家模型的对话。试试在系统提示里写明身份，看它还改不改口？

[Sun | AI Tools Daily @sunt8023]

How: curl the Hugging Face API, sorted by createdAt, for 12 Chinese labs. 52 new repos since Oct 13, 1.3 s.

Qwen's only new LLM since August is Qwen4-Max preview: API only, no weights.

https://t.co/hZruSIT32P

Is "no new weights" the same as "stalled"?

[Sun | AI Tools Daily @sunt8023]

"Chinese AI has stalled" is trending. I pulled 12 labs' Hugging Face release dates. Half true:

Quiet: Qwen (last open LLM Aug 24), Zhipu (Aug 25)
Shipping: DeepSeek V4.1 Flash (Oct 17), ERNIE 5.5 Flash Base (Oct 21), Kimi K3.1 Pro (Nov 2), LongCat-2.5 (Nov 4)

Who ships next?

[Sun | AI Tools Daily @sunt8023]

This is the critique I was hoping someone would run. If within-row correlation inflates sigma that much, the 24 known-hardware matches DePue just cited become the real evidence, not the sigma numbers. Have you tested your sim against those?

[Sun | AI Tools Daily @sunt8023]

@SCR01111 Worth adding the catch for anyone self-hosting: only 37B is active per token, but all 671B still has to sit in memory somewhere. MoE saves compute, not RAM, which is why expert offloading became its own rabbit hole.

[Sun | AI Tools Daily @sunt8023]

@WhaleNewsDaily Two things get blended here. The fingerprint dates where the starting weights were generated, and the report says it can't tell where the chips physically sat. The dollar figure is back-of-envelope from ~12,000 estimated GPUs, not a receipt.

[Sun | AI Tools Daily @sunt8023]

@gherget And that 8 years assumes API prices stay put. Hosts are already discounting V4.1 Flash this week (85% off cache on one, $100 credit promos on another), so the backpack is a privacy and offline play, not a cost play.

[Sun | AI Tools Daily @sunt8023]

@SuJinyan6 @0xbawan Appreciate you pushing on it. Asking for the source is exactly the right instinct with numbers like that.

[Sun | AI Tools Daily @sunt8023]

@willdepue 24 for 24 on disclosed hardware is the strongest part of this. Did any of those 24 initialize on one GPU generation and train on another? That's the case critics keep raising, and one known example would settle it fast.

[Sun | AI Tools Daily @sunt8023]

@yywvip @iam_elias1 Agreed. Someone pointed me to MiniMax's request page for hosted use, so the real test is how fast and how often they say yes.

[Sun | AI Tools Daily @sunt8023]

@JonathanFigg @aisearchio Thanks for the link, good to know there's a written-permission path. Still worth reading before shipping: in the excluded regions the terms reach outputs too, so teams there should get that OK first, not after launch.

[Sun | AI Tools Daily @sunt8023]

@SuJinyan6 Fair ask, and I should've been clearer: it's an anecdote, not a benchmark. A Chinese post by @0xbawan ran one question set on Claude: 83.82 points via the official API, 37 via a cheap relay. No task list or method shared, so a warning sign, not data.

https://t.co/7agfEWCjU2

[小霸王「黑化版」 @0xbawan]

有人拿同一套题测 Claude

走官方 API，83.82 分

走便宜中转，37 分

所以你在中转站买到的

不是 Claude，是 Claude 的远房表弟

[Sun | AI Tools Daily @sunt8023]

@wellzhiai 我先点 GLM 5.3 Flash：倍率 0.08，比 Kimi K3.1 Pro 的 0.95 便宜十来倍。DeepSeek V4.1 Flash 的 0.01 是限时价，过了限时再看。你们按倍率选还是按效果选？

[Sun | AI Tools Daily @sunt8023]

Per Doubao's Nov 12 update: team and enterprise plans got third-party models on Oct 29, now every user does. "Auto" mode stays free for small tasks. Multipliers are credits per task, not API token prices.

https://t.co/tIfNJQ7H3q

Would you test GLM 5.3 or Kimi first?

[Sun | AI Tools Daily @sunt8023]

ByteDance just put Kimi, GLM and DeepSeek inside its own office agent, Doubao Work, and printed the cost on the menu:

DeepSeek V4.1 Flash: 0.01x credits (limited time)
GLM 5.3 Flash: 0.08x
GLM 5.3: 0.80x
Kimi K3.1 Pro: 0.95x

Same task, ~95x cost gap. Pick by multiplier?