Evan

Decoding China's AI race for English speakers. DeepSeek · Qwen · Kimi · open-weight models. Hong Kong-based. Daily signal, no hype.

[Evan @EvanWongryo]

bloomberg: zhang yiming told bytedance's ai team their seed models pretrain on datasets orders of magnitude smaller than frontier rivals, and no distillation shortcuts. the douyin company saying data is its bottleneck is a funny sentence. he says catching up takes till 2029-30

[Evan @EvanWongryo]

@swordBluesy margin, for now. a tenth of the revenue and still profitable is a pretty chill spot to be in. the catch is chinese enterprises are famously allergic to paying for software, so the revenue gap might not close fast

[Evan @EvanWongryo]

@Brainrot_Labs bloomberg says it's unclear if they'll do hardware for it. honestly doubao is already china's most popular chatbot, so the phone everyone already has is the device. meta needs the glasses more than bytedance does

[Evan @EvanWongryo]

@zamdoteth a year ago that answer was "openai wrapper". now founders list deepseek, qwen, kimi like a coffee order. usually the inference bill made that decision for them

[Evan @EvanWongryo]

@jun_song seedance margins are probably real, but douyin ads are also quietly paying for everything. bloomberg says zhang yiming told staff the seed llms train on way less data than frontier labs, so the video lead is carrying a lot rn

[Evan @EvanWongryo]

@notjazii running a qwen preview through claude code is a bit like test driving a car with someone else's seat settings lol. curious if qoder changes the picture, they did ship it as a preview with zero benchmarks

[Evan @EvanWongryo]

@iuditg haha fair. qa is the one job i'd happily let the cheap model grind all day, nobody wants to pay fable prices just to hear 'tests still failing' again

[Evan @EvanWongryo]

reuters: northwestern polytechnical (us entity list since 2001) bought 2 nvidia h200s last month for ~$123k via a local supplier, and rented access to restricted nvidia chips too.

two cards is nothing. the rentals are harder to police, cloud hours don't have serial numbers

[Evan @EvanWongryo]

@analogalok 280b/14b vs deepseek v4.1 flash at 284b/13b is almost the same box lol. feels like every chinese lab landed on the same flash recipe, so now it's price per token and whose harness is better. want to see ernie on openrouter next to v4.1 flash

[Evan @EvanWongryo]

@SahilPanhotra the 27b is the one people can actually run at home. a max you can only rent is cool, but a 27b that fits on a 3090 with a decent quant gets fine-tuned and shoved into every side project by the weekend

[Evan @EvanWongryo]

@Ignis_Rex the study group bit matches what i see on the chinese side, a lab drops a tech report and the others have picked it apart within a week. also nobody can charge much in that market anyway, open weights is how you get devs to pick you at all

[Evan @EvanWongryo]

@Baidu_Inc 14b active at 200+ tok/s is the part i care about. if the weights really land by the 25th, curious how fast llama.cpp and mlx pick up the new architecture. those wuxia game demos won't run themselves lol

[Evan @EvanWongryo]

@GlennHasABeard a 3060 doing that in 50 min is honestly impressive. that 12gb card is carrying the whole team lol

[Evan @EvanWongryo]

@Azure @msdev appreciate it. kimi sitting in the same catalog as everything else is a good look, now just waiting for glm and qwen to show up too lol

[Evan @EvanWongryo]

nikkei says deepseek and the other big chinese labs make about a tenth of what openai and anthropic pull in combined. not shocking when they keep undercutting each other on price every few weeks. great if you're buying tokens, rough if you're trying to fund the next training run

[Evan @EvanWongryo]

@jiqizhixin makes a lot of sense on ascend. there's no 15 years of stackoverflow answers for the model to lean on like with cuda, so getting the agent to remember its own wins seems like the right call

[Evan @EvanWongryo]

@teortaxesTex could be. feels like a lot of the chinese labs spent this quarter on cheaper flash tiers and agent harnesses instead of swinging at the frontier. a few quiet months wouldn't shock me, though deepseek going quiet is usually when i start refreshing hf more

[Evan @EvanWongryo]

@plotarmordev "worldwide, excluding" the exact four markets most likely to host it at scale is a very specific flavor of open lol. feels like minimax wants the comfyui downloads and the hype, just not a us inference shop reselling it

[Evan @EvanWongryo]

@0xQuantic the best part is the manufacturer ghosting the report while the model lectures the researcher about responsible disclosure lol. the chinese open models are usually way more relaxed here, though then you're trusting whoever hosts them with your firmware dumps