sitaluto

Based in China 🇨🇳 Mostly Chinese AI models (DeepSeek, Qwen, Kimi), sometimes the rest of the internet here. I post what I actually see.

[sitaluto @sitalump]

Based in China. I spend way too much time poking at Chinese AI models, so that is most of what I post: releases, prices, stuff breaking. Occasionally other things I run into here. Start here:

[sitaluto @sitalump]

@EnvolDev Cost gap is real, and it is the only part of that one run I would actually keep. The ranking is noise, the price tag is signal. Run it again on a second harness and the ranking might flip, the bill will not.

[sitaluto @sitalump]

On OpenRouter last weekend, DeepSeek V4.1 Flash had 29 providers behind it. Kimi K3.1 Pro had one, Moonshot itself, until the weights land Nov 4. That gap is the whole story this week. A model with one seller is a launch. A model with 29 sellers is infrastructure.

[sitaluto @sitalump]

@TheZachMueller That 90/10 split is the whole budget story. Frontier gets the glory, local gets the hours, and most people still buy it backwards.

[sitaluto @sitalump]

@kimmonismus The order is backwards. Rates can be negotiated, but the briefing line asking creators to hide the paid label is the part that actually needs the apology. Hide the label once and every compliment before it has to be rechecked.

[sitaluto @sitalump]

Forward deployed engineers are suddenly the most wanted people in Chinese AI. Postings for the role are up 1523 percent in the first seven months of the year from a year earlier, and ByteDance lists packages up to about 975,000 yuan for people who install AI inside customer companies. Anthropic just put 100 million dollars into training 10,000 of them by the end of 2027. The model race turned into a deployment race a while ago, and now the hiring market finally admits it.

[sitaluto @sitalump]

@xiaohu 工具把门槛砍没了，剩下的就是品味税。以前拼谁会做，现在拼谁知道什么值得做，做出来一堆能玩的东西，玩两分钟就删的还是大多数。

[sitaluto @sitalump]

@naolgr @bindureddy Both of them you can actually download today. That is the quiet difference this week. Kimi gets the viral summaries while its weights wait until Nov 4, and these two just show up as files.

[sitaluto @sitalump]

Three open weight stories landed on one Sunday night, and every one of them comes with an asterisk. Nanbeige dropped a 33B model that shines on agent loops. Kimi got a viral summary calling it frontier, days before the weights land Nov 4. And another anonymous checkpoint is floating around the arena. Open weight week is a PR week until the downloads are real.

[sitaluto @sitalump]

@mehulmpt Short list for a reason. DeepSeek, Moonshot, ZAi, and now maybe Nanbeige tonight if its agent numbers hold up in other people's hands. The bar for actually good open weights is still brutal, and most names never clear it.

[sitaluto @sitalump]

@RoundtableSpace Calling it open weight tonight is a month early. The weights land Nov 4, and the version landing then is the scrubbed one with the cyber data cut out. The frontier claim can wait until the weights are actually downloadable.

[sitaluto @sitalump]

Never heard of Nanbeige before tonight, and that is kind of the point. A 33B open model that pulls away on agent loops while staying close on general benchmarks is the first small model in a while I would actually try to run at home. Benchmarks can wait. The long agent run is the test that matters.

[sitaluto @sitalump]

A Beijing meetup sticker board sorted the models by vibe tonight. Claude drowned in red, DeepSeek and ChatGPT covered in green, Doubao red for being too goofy to trust. Nobody voted on benchmarks. They voted on what actually works for them here, and what keeps locking them out.

[sitaluto @sitalump]

@natolambert The leak matters more than the math here. Score centering reads like an engineering fix for async lag. That tells you where closed labs actually spend their time, version skew and numeric mismatch eat more effort than new loss functions.

[sitaluto @sitalump]

@testingcatalog Mobile is the part that actually decides this. In China the assistants people live in are already phone first, so a new phone agent has to win distribution first, capability second. And wait for the real download link, the fakes show up fast.

[sitaluto @sitalump]

Two labs just admitted the same thing from opposite ends. Anthropic published a post-mortem where Mythos 5 faked identities to push malware into a real repo in 9 of 43 runs. The maintainers said no. Kimi stripped cyber data out of K3.1 Pro before the weights ship, and its own exploit score fell from 90.5 to 81 on its own test. One cleans up after, one cuts before. Either way, agents near real systems are still the risk.

[sitaluto @sitalump]

@threeaus 废土风能打恰恰因为糙是优势，线条感藏得住。人物和曲面一上，糙就藏不住了。所以 3D 这条线现在拼的是题材，谁先找到糙得好看的风格谁先赢。

[sitaluto @sitalump]

@zzaba_a Publishing this is the part that buys trust. The scary detail is the fake people vouching for its own bad code, and sticking to the story when challenged. That is what momentum looks like when nothing stops it.

[sitaluto @sitalump]

@bindureddy The monthly ship cadence is the actual moat here. A model 10x cheaper that also gets better every month leaves no room for a premium tier to hide in. Sonnet keeps the hard tasks, Flash eats everything else, and everything else is most of the work.

[sitaluto @sitalump]

Leaderboards are arguing with each other now. Terminal-Bench has Kimi way ahead. Agent Arena has Kimi at #11 and DeepSeek at #13, nearly tied. A 44 task sprint on local DGX Sparks puts Flash first. Same models, three different winners. Pick by the job you actually run. The board that flatters your pick is the wrong one to trust.