First Week of August: Two Chinese Giants Drop Heavy Models
Alibaba pushed Qwen3.8-Max to 2.4 trillion parameters. A week earlier, Moonshot AI had already dropped the full 2.8-trillion-parameter Kimi K3 weights onto Hugging Face.
One is "about to open-source," the other "already did."
But the question a small-business owner should actually ask is not "which is stronger" — it is "can I actually afford it and run it."
Let's Talk Money First
Qwen3.8-Max (Alibaba, released August 3)
- 2.4T sparse MoE, 95B active parameters per token, 1M-token context, with vision
- API: 12 yuan input / 36 yuan output per million tokens in China; $2 input / $6 output overseas, $0.25 cached
- Priced at just 40% (input) and 24% (output) of Anthropic Opus 5
- Max weights open-source next week; a 27B variant also goes open
Kimi K3 (Moonshot AI, weights open July 27)
- 2.8T parameters, 104B active (16 of 896 experts), 1M-token context, native vision
- API: $3 input / $15 output per million tokens cache-miss, $0.30 cached
- Released under a custom Kimi K3 License (not MIT — read it before commercial use)
- Quantized weights still weigh about 1.4 TB
Who Is Stronger? Don't Just Count Parameters
Qwen3.8-Max ranks 5th in Arena Text, 2nd in Vision, 4th in CodeArena. On PaperBench it scored 93.0, beating Fable 5's 88.8; on instruction-following IFBench it hit 82.8, well ahead of Fable 5's 63.5.
Kimi K3 scored 42.0 on the long-horizon SWE Marathon — triple GPT 5.5 — and posted 91.2 on BrowseComp and 95.0 on DeepSearchQA. Solid on knowledge work.
Plainly: Qwen3.8 is more balanced across office and vision; Kimi K3 has a higher ceiling on long-agent and knowledge tasks.
The Real Verdict: Open Does Not Mean You Can Run It
For 99% of small businesses, "open weights" is a comfort prize. In practice, the API is what you actually use.
Both APIs are absurdly cheap. Qwen3.8 at $2/$6 overseas might cost you tens of dollars a month for a support bot. Kimi K3 at $3/$15 is not expensive either.
Self-hosting is another story. Kimi K3's 1.4 TB weights need at least an 8x H100-class setup. Qwen3.8-Max is just as massive — open-sourced or not, you probably cannot run it. Private deployment of the full model is a tens-of-millions investment.
A Signal: Chinese Models Are Going Global
This is not a closed-door exercise. Per OpenRouter, as of July 27 Chinese LLM weekly usage was 14.1x that of US models — 13 weeks atop the chart.
The proof is in deployment: Qwen3.8 entered "deep testing" for Tesla's in-car systems; Coinbase set Kimi K2.7 (K3's predecessor) and Zhipu GLM-5.2 as default engineer tools, cutting AI cost nearly in half via smart routing.
Plainly, price-performance is becoming the hard currency of Chinese models. When Silicon Valley shops switch for being "cheap and capable," the game's underlying logic has shifted.
Why Alibaba Agreed to Open-Source Max
Qwen-Max-class models never went open before. The calculus is clear: lock in long-term private-deployment customers with weights, pull developers into the ecosystem with cheap APIs.
Alibaba Cloud's AI revenue has grown triple digits for 11 straight quarters; token demand is up 10x in six months. Open-sourcing is not charity — it accelerates token monetization. Let you use it free, and you pay once you are hooked.
Which Should You Pick
- Need vision + office + fast integration on a compliant domestic API → Qwen3.8, brutal price-performance, and you can experiment with private deployment once weights drop
- Doing frontier research, long-context knowledge bases, agents, and not afraid of a custom license → Kimi K3, genuinely open weights, higher ceiling
- Want to buy GPUs and run the full model yourself → both will talk you out of it; use the API, or wait for the 27B small variant
Why This Matters to You
You run a Taobao store, do foreign trade, or manage a team of a dozen. Do not let "trillion parameters" dazzle you.
Do the math: Qwen3.8's API handling 1,000 support queries a day might cost you a couple hundred yuan a month. Self-hosting full Kimi K3 is a tens-of-millions outlay plus an ops team.
For small business, "calling" always beats "owning." Get the API working and see real ROI before anything else.
Bottom Line
This wave of Chinese models is not about whose spec sheet scares you more — it is about who lets you use it affordably and reliably. Qwen3.8 and Kimi K3 have pushed the floor to a historic low. Now it is about who ships the use case first.