Chinese AI models in October 2026: state of play

This is a snapshot of five Chinese AI labs as of early October 2026: DeepSeek, Alibaba’s Qwen, Moonshot’s Kimi, Z.ai (Zhipu) and MiniMax. I used each vendor’s own pages for dates, licenses and prices, and every benchmark claim below is the vendor’s, not mine. I did not research ByteDance, Tencent or other labs, so this is not a full map.

What shipped

Prices are US dollars per million tokens, read from the vendors’ official pages on 5 October 2026.

Vendor	Latest notable release	Weights and license	API price, input / output
DeepSeek	V4.1-Flash, 10 Sep 2026 (V4-Pro went GA on 13 Aug)	MIT on V4.1-Flash	Flash: $0.15 / $0.60 off-peak, $0.30 / $1.20 peak. V4-Pro: $0.66 / $1.98 off-peak
Qwen	Qwen3.8-Max, 2 Aug 2026	Apache 2.0; 27B dense and 2.4T-A95B weights reported 14 Aug	$2 / $6 (international)
Kimi	K3, API on 16 Jul 2026, weights on 27 Jul	Kimi K3 License	$3.00 / $15.00 (cache miss)
Z.ai	GLM-5.3, 14 Aug 2026, weights on 28 Aug	GLM-5.3 license (GLM-5.2 was MIT)	$1.4 / $4.4
MiniMax	M3, 1 Jun 2026	MiniMax Community License	$0.30 / $1.20 up to 512K input

Input prices are cache-miss. DeepSeek’s peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays; every other hour is off-peak at half the price. Its 10 September note said deepseek-v4-pro requests would route to V4.1-Flash from 14 September, yet in early October the pricing page still lists V4-Pro. I would treat that lineup as unsettled.

“Open” means different things

Open weights is not one license. Reading the actual text matters more than the label.

• DeepSeek V4.1-Flash and Qwen3.8: standard permissive licenses (MIT and Apache 2.0 respectively).
• Kimi K3: The grant is broad, but a company running a “Model as a Service” business that, with affiliates, passes $20 million in revenue over any 12 months needs a separate agreement. Products above 100 million monthly users or $20 million monthly revenue must display “Kimi K3”. Internal use is exempt.
• GLM-5.3: Z.ai replaced the MIT license used for GLM-5.2. Hosts with over $10 billion in revenue “must pass Z.AI’s security review” before commercial use. The company says it held the weights for two weeks of safety evaluation.
• MiniMax M3: The grant covers non-commercial use. Commercial use requires showing “Built with MiniMax M3” and sending a one-time notice; above $20 million yearly revenue you need written authorization. Fine-tuned deployments for commercial purposes count as commercial use.

What the vendors claim

• DeepSeek says V4.1-Flash’s new pre-training and larger-scale RL “deliver benchmark results ahead of flagship models, including DeepSeek-V4-Pro.”
• Qwen says the 27B model outperforms the larger Qwen3.7-Plus in coding and office tasks.
• Moonshot says K3 shows frontier-level performance across its suite, while its overall performance “still trails” Claude Fable 5 and GPT 5.6 Sol. Its own limitations list also notes a gap in user experience against those two.
• Z.ai calls GLM-5.3 “the most capable open-weights model for coding,” and also writes that it “remains behind Claude Fable 5” on its in-house coding bench.
• MiniMax reports 59.0% on SWE-Bench Pro and says M3 is “the first and only open-weight model” to combine frontier coding and agentic work, 1M context and native multimodality.

Moonshot and Z.ai both name a gap to the closed models. I find that more informative than the leaderboard numbers, which each vendor ran on its own setup.

Adoption among Western developers

CNBC reported on 26 September, using data from OpenRouter and Vercel, that Chinese models made up 57%-67% of OpenRouter tokens in the week of 14 September, up from 6%-13% in February. On Vercel the share rose to 55% in August from 11% in January.

Two caveats from the same article. Tokens are not spend: it notes that “U.S. frontier models still attract more overall spending.” And about half of OpenRouter tokens come from US companies. An earlier CNBC piece (7 July) put US companies’ share of tokens on Chinese models above 30% every week since 8 February.

Vercel’s Harpreet Arora gave the reason as price: “Once a model meets the quality bar for the job, that price difference becomes compelling.”

Chips and export controls

The US restricts Chinese AI companies from buying the most advanced chips. CNBC reported on 19 August that Chinese firms had reportedly accessed Nvidia computing power overseas, and that lawmakers were weighing closing that cloud-access gap. In April, Reuters reported (via Capacity) that ByteDance, Tencent and Alibaba were seeking Huawei Ascend 950 orders after DeepSeek optimised V4 for domestic silicon. Two US House committees are also investigating the rise in Chinese-model use. I cannot tell from these sources how much of any model was trained on which chips.

What it means for a small developer

This part is my analysis.

Cost. The official list prices for DeepSeek Flash, GLM-5.3 and MiniMax M3 sit at or below $1.5 per million input tokens; Qwen3.8-Max and Kimi K3 are priced higher. Prices also move: DeepSeek changed its pricing on 16 August and again on 10 September. I would compare cost per finished task on my own workload, not per token.

Self-hosting. K3 weights are roughly 1.5 TB and Moonshot recommends 64 or more accelerators, so that is not a solo-developer option. Of the sizes I checked (K3, M3, V4.1-Flash, Qwen3.8), only the 27B Qwen under Apache 2.0 is sized for a single machine, and I would still check its memory needs before planning around it.

Licenses. The revenue thresholds ($10 billion, $20 million) are far from a small product. MiniMax M3 is the exception that matters at small scale, because commercial use triggers attribution and a notice. Read the LICENSE file of the exact version you download; Z.ai changed terms between two point releases.

Data and compliance. DeepSeek’s privacy policy says it stores personal data in the People’s Republic of China. Calling a Chinese vendor’s official API sends your users’ prompts to that infrastructure, which may be a problem for client contracts or regulated data. Running open weights through a third-party host is a different path: M3, for example, is listed on Hugging Face with providers such as Together AI, Fireworks and DeepInfra. That host’s data terms then apply, so I would read them. I would also keep the model name in config, so switching costs one line when prices, licenses or client requirements change.