From our other sites
The story
Hugging Face has published download statistics for January–July 2026, revealing that the Chinese open model Qwen racked up nearly twice as many downloads as Gemma. The thread saw a split of opinions over the technical standing of Chinese-made AI, alongside a lively exchange of real-world stories about switching to local LLMs and tips for managing subscriptions.
Qwen, Kimi, MiniMax… the momentum behind Chinese AI models these days shows no sign of stopping. Statistics on open models released by Hugging Face on August 14 (US time) reveal that Chinese AI models are continuing to expand their share.
This report comes from Hugging Face, where a huge number of AI models are uploaded, and covers data from January 2026 through July 31. Among the notable trends observed during this period is the following.
The rise in frontier models
Source: pc.watch.impress.co.jp / Original article here
What people said
For chat, Gemma's Japanese is better.
For coding, Qwen's the way to go.
That's exactly how it feels using them, no joke.
Hold on now.
Qwen's been around way longer than Gemma — since April 2023, right from the start.
It's been the king of open models since early on, right after Meta's Llama in February 2023.
Gemma didn't show up until a year later, in April 2024.
If anything, the story should be about how much Gemma has caught up.
"The rise of China" this, "China's latest thing" that —
Impress Watch has gotten pretty dumb too, huh.
For PC or programming questions, even a 9B model can hold a conversation at a level where you don't need the internet, but LLMs just can't handle current-events questions.
Ask a local LLM about the latest news and it treats it as fake and just refuses to answer at all.
Re: #4
Some "king" — the moment they went closed, they lost consumer interest instantly and had to scramble back to the open-source route.
Must be nice doing work where it's fine if your data gets siphoned off.
A publicly listed company couldn't possibly place an order with them.
Who's using an LLM to ask that kind of question anyway lol
Who do you think is the one serving Gemma in the first place?
You don't even understand what "open model" means,
and you're out here saying "you'd better…"
Even on the $200 plan, Fable 5 would hit a week's worth of rate limits after running for just a single day — totally unusable.
So starting this month I switched to Kimi K3's $99 plan instead, and the performance is basically on par with Fable 5, except now I don't have to worry about rate limits.
And as of yesterday I've got Qwen 3.8 27B running locally, handing it lighter tasks like coding.
Qwen 3.8 27B performs about the same as Opus 4.8, so I've got nothing to complain about.
Honestly, I just want to say thank you, China.
Thanks for the report. Gonna copy what you're doing.
Read this and tried it on my own machine (Ryzen AI Pro 370), but Qwen 3.8 27B only managed 6 tokens/s even at Q3_K_M lol
Memory bandwidth really is everything.
Guess I'll quietly go back to Ornith 1.0-9B.
Claiming the same performance as a frontier model when the parameter count is nowhere close is a stretch to begin with.
Parameter count is about the amount of knowledge, not a ceiling on reasoning ability like you'd need for coding.
Saying reasoning doesn't come from knowledge — typical of Japan, land of the flash-quiz trivia shows. (a jab at Japan's love of quick-recall trivia/quiz TV culture)
"Parameter count is the amount of knowledge" is completely wrong. Go read a textbook.
Like a lot of people have pointed out, no matter how cheap the pay-as-you-go token price gets, it's hard to beat subscription plans on cost, OpenAI included.
I'd get it if the issue were licensing restrictions, but…
There's also the question of what you do once the subscription's gone.
Uh, you do know that using the AI agent API through Claude Code is billed sep-a-rate-ly, right?
There's no "rise" when it's just benchmarks with no real track record.
State-of-the-Art Coding Agents:
Available in 9B-Dense, 31B-Dense, 35B-MoE, and 397B-MoE
(post-trained on top of Gemma 4 and Qwen 3.5),
achieving state-of-the-art performance among open-source models of comparable size on coding benchmarks such as Terminal-Bench 2.1, SWE-Bench, NL2Repo and OpenClaw.
https://huggingface.co/ornith-ai/Ornith-1.0-9B
Just on Hugging Face alone this has 2.4 million downloads, so that's pretty impressive.
Even running an agent loop, it doesn't beat that number. How many sessions are you running?
Wait, are you running it single-threaded?
What a waste — people are apparently running 100 or 200 in parallel nonstop.
Would be a shame to build a dual-GPU rig and have it turn out underwhelming.
So I want to try it on an overseas GPU cloud, but I'm scared to enter my credit card info.
And apparently you need to charge up something like 10,000 yen (~$65) before payment even goes through, which is even scarier.
By the way, you can pick GeForce-series GPUs too, but using those in a data center is against the license terms (except for blockchain use), so any provider offering that is sketchy.
I thought Sakura Internet's GPU servers would be cheap and safe, but then there was that personal data leak thing…
So, guess I'll just rent an AWS GPU instance to test it out.
Quit yapping and just go do it already.
That level of worry was a 2023 thing.
You're three years behind too.
In this context it just means LLMs. Adding "local" would be wrong.
Guess we're an AI backwater now, better go beg China for help lol
Japan's being left behind?
Then which country do you think Gemma is even from?
I'd love to hear your answer.
Background and Key Points
Hugging Face’s usage stats are one of the few public proxies for how widely an open-weight model is actually being run, since neither Alibaba (Qwen) nor Google (Gemma) publish adoption numbers directly. Qwen has been part of the open-model landscape since shortly after Meta’s Llama kicked off the current wave in 2023, while Gemma only entered in April 2024, so “Qwen overtakes Gemma” is less a sudden coup than a gap that predates this report. For Japanese readers, the relevant backdrop is that Chinese labs (Alibaba’s Qwen, Moonshot’s Kimi, MiniMax) have leaned hard into releasing permissively-licensed weights rather than closed APIs, which is exactly the format hobbyists and small companies can download and run on their own hardware — hence the local-LLM enthusiasm running through the thread.
What actually divided the thread wasn’t China-vs-Japan so much as whether raw download counts mean anything: several posters objected to the “rise of China” framing itself, arguing Qwen’s lead is old news and that benchmark wins without a real-world track record (#41) don’t equal quality. A separate, more heated split was over whether a 9B–35B local model can genuinely substitute for a frontier subscription model like Fable 5 or Opus, with posters disagreeing over whether parameter count maps to reasoning ability at all.
One thing the thread never engages with: the “Ornith” model posters cite as a homegrown success (#48–49) is itself post-trained on top of both Gemma 4 and Qwen 3.5 — so the China/Japan-Google rivalry framing obscures that the leading downloads are hybrids built on both lineages, not a clean national contest.
*This article is composed as an excerpt and summary of the 5ch (Business News+) thread “[AI] Qwen Nearly Doubles Gemma — The Rise of Chinese AI as Seen in Hugging Face Statistics.”
Leave a Reply