- 29
- September
"China AI Update, September 2026: DeepSeek V4.1, Qwen 4 Imminent, Anthropic's Kimi Accusation, Huawei's Chip Sprint, and Open Models Within 10 Points of the Frontier" — the short answer is in September 2026 a Chinese open-weight model became the highest-scoring open model in the world, 7–9 points behind the top closed models, and Chinese models held 7 of the 10 most-used slots, while in the same month Anthropic accused seven Chinese labs of harvesting 190 million Claude responses and China's regulator summoned all seven. This article runs from Kimi K3 in July to the Apsara Conference and Huawei Connect in late September, covering models, prices, chips, money and rules, and closes with what Thai organizations should do. It follows our August 2026 AI roundup.
In one line: China AI in September 2026: DeepSeek V4.1-Flash on a new architecture at lower prices · Xiaomi MiMo-V2.6-Pro is the top open-weight model worldwide (46 vs 53 for the best closed models) · Alibaba previews Qwen 4 with its Zhenwu V900 chip and a 20 GW data-center target · Huawei pulls Ascend 960DT to early 2027 · Zhipu raises $5 billion · Anthropic accuses seven Chinese labs of harvesting 190 million Claude responses and the CAC summons all seven.
Before you read: Everything here was checked against each company's own announcements, official API documentation, Anthropic's published report, and international news coverage as of 29 September 2026. Prices are in US dollars or Chinese yuan as the vendors publish them, with no conversion to baht. Benchmark figures are the vendor's own or Artificial Analysis's, as stated at each point. The distillation allegations are Anthropic's alone; the accused companies had not responded at the time of writing.
September 2026 brought five big Chinese AI stories at once. New models from DeepSeek and Xiaomi took the top spot among open-weight models worldwide. Alibaba previewed Qwen 4 alongside its own chips. Huawei pulled its Ascend chip roadmap forward by three quarters. Zhipu raised another $5 billion after quadrupling revenue. And Anthropic published a report accusing seven Chinese labs of quietly harvesting 190 million Claude responses, which ended with China's own regulator summoning all seven companies. This article puts the events in order, explains what each one means, and closes with what Thai organizations should do about them.
Timeline: What Happened on the Chinese Side Since Mid-2026
| Date | Event |
|---|---|
| 16 Jul | Moonshot releases Kimi K3, 2.8 trillion parameters with a 1M-token context, the largest open-weight model to date · weights published 27 Jul |
| 29 Jul | Moonshot closes a $3.5 billion round at a $35 billion valuation |
| 3 Aug | DeepSeek takes #1 by token volume on OpenRouter for the first time, ending Google's 51-week lead |
| 7 Aug | Reuters reports Alibaba will require revenue-sharing agreements from large commercial users of its next open-weight Qwen release |
| 13 Aug | DeepSeek V4-Pro reaches general availability and introduces peak / off-peak pricing, off-peak at half price (effective 16 Aug) |
| 14 Aug | Zhipu releases GLM-5.3 · Hugging Face reports Qwen downloaded 3 billion times in six months |
| 28 Aug | Alibaba releases Qwen3.8-Flash-Next, an experimental model built on the Qwen 4 architecture |
| 31 Aug | Zhipu reports H1 results, revenue up nearly 400%, and names GLM-6.0 as its next generation |
| 2 Sep | CAC publishes second-phase results of its 2026 campaign against AI misuse |
| 4 Sep | Alibaba Cloud holds Qwen Conference Thailand 2026 in Bangkok with nearly 400 attendees |
| 10 Sep | Anthropic publishes a report accusing seven Chinese labs of illicit distillation totalling 190 million exchanges · DeepSeek releases V4.1-Flash on a new architecture and cuts API prices |
| 12–16 Sep | Zhipu raises about $5 billion through a share placement and zero-coupon convertible bonds |
| 17–18 Sep | Huawei Connect 2026 in Shanghai: Ascend 960DT pulled forward to Q1 2027, three quarters early |
| 20 Sep | Alibaba releases Qwen-Image-2.1, a 7-billion-parameter image model that runs on consumer GPUs |
| 22 Sep | Alibaba's Apsara Conference: Qwen 4 in training, the Zhenwu V900 chip, and a 20 GW data-center target · Xiaomi open-sources MiMo-V2.6, the top-ranked open-weight model · CAC summons the seven companies named in Anthropic's report · ByteDance releases Doubao-Seed-Translation for 28 languages · Anthropic releases Claude Opus 5.5 |
| 23 Sep | Alibaba Cloud announces data-center expansion in eight countries, reaching 107 zones in 31 regions |
| 28 Sep | Anthropic releases Claude Sonnet 5.5 with classifiers that block reasoning extraction |
Sources: company announcements, DeepSeek API documentation, the State of Open Source AI v1.1 report, and reporting by Bloomberg, Fortune, CNBC, TechNode and The Next Web (29 Sep 2026)
1. New Models: DeepSeek Changes Architecture, Xiaomi Takes #1, and Qwen 4 Is in the Oven
This month's Chinese releases were not just newer versions. Two labs changed architecture, and one lab few people watch, Xiaomi, became the highest-scoring open-weight model in the world on the Artificial Analysis index.
| Model | Company | Date | What changed | API price (per MTok) | Open weights |
|---|---|---|---|---|---|
| DeepSeek V4.1-Flash | DeepSeek | 10 Sep | The smallest model in what DeepSeek calls a "new architecture family" designed for "a higher capability ceiling, faster inference, higher throughput, and scaling to larger models" · native image understanding · 1M context, 384K max output | $0.30 input / $1.20 output at peak · half price off-peak · cache hit $0.006 | Not stated |
| DeepSeek V4-Pro | DeepSeek | 13 Aug (GA) | Responses API support and three effort levels (low / high / max) · no vision yet | $1.32 input / $3.96 output at peak · half price off-peak | Yes (April preview) |
| MiMo-V2.6 Pro / Flash | Xiaomi | 22 Sep | Trillion-parameter Pro model taking text, image, video and audio · Artificial Analysis score 46, the highest of any open-weight model (Kimi K3 44, GLM-5.3 45, top closed models 53) · Xiaomi puts the RL stage cost at $2.62 million for Pro and $850,000 for Flash | Not compiled | Yes |
| Qwen3.8-Flash-Next | Alibaba | 28 Aug | An experimental model Alibaba openly calls "a preview of the architecture" behind Qwen 4 | — | Yes |
| Qwen-Image-2.1 | Alibaba | 20 Sep | 7-billion-parameter image generation and editing that runs on a consumer GPU | — | Yes |
| GLM-5.3 | Zhipu (Z.ai) | 14 Aug | Improved post-training · GLM-5.3 Flash processed 60 trillion tokens in its first six days | API prices raised about 101% in H1 (per the earnings call) | Yes |
| Kimi K3 | Moonshot | 16 Jul | 2.8 trillion parameters, 1M context, the largest open-weight model · debuted at #3 worldwide on Artificial Analysis | $3 input / $15 output | Yes (custom license) |
| Doubao-Seed-Translation | ByteDance | 22 Sep | Translation model covering 28 languages on Volcano Engine | — | No |
Sources: DeepSeek Pricing and Change Log pages (29 Sep 2026), TechNode 22 Sep, Xiaomi's announcement, Fortune 16 Jul, Zhipu H1 earnings call 31 Aug · DeepSeek's "peak" window is 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday
Three things to take from the table. First, DeepSeek changed how it prices, from one flat rate to time-of-day rates with off-peak at half price. Peak hours are morning to late morning in China, which lines up with the Thai working day. A Thai organization running DeepSeek during office hours pays full price; overnight batch jobs get the half rate. Second, Xiaomi, which most people think of as a phone maker, owned the top open-weight score as of 22 September, ahead of Moonshot and Zhipu, who got there first. Third, Alibaba has not shipped Qwen 4. It says the model is in training and coming "very soon," and that Qwen 4.5 and Qwen 5 will scale to 5–10 trillion parameters. Anyone planning to move workloads to Qwen should wait for version 4.
2. Chinese Open Models Trail the Frontier by Under 10 Points but Already Own the Volume
The number that best describes the state of play is not a test score. It is how many tokens people actually run. The September 2026 State of Open Source AI report, using OpenRouter data, finds that eight of the ten highest-volume models are open weights, and seven of those eight were built in China. Chinese models peaked at 46% of routed tokens, and DeepSeek has more than 26,000 enterprise accounts.
| Indicator | Figure | Source / date |
|---|---|---|
| Artificial Analysis Intelligence Index, top closed models (Claude Fable 5.1, GPT-6 Astra) | 53 | Artificial Analysis, Sep 2026 |
| Highest Chinese open-weight models: MiMo-V2.6-Pro / GLM-5.3 / Kimi K3 | 46 / 45 / 44 | Artificial Analysis, 22 Sep 2026 |
| Expert knowledge-work gap (GDPval-AA), Fable 5 over Kimi K3 | 92 Elo points | State of Open Source AI v1.1 |
| Top 10 models by token volume on OpenRouter | 8 open · 7 Chinese | OpenRouter, Aug 2026 |
| Qwen downloads in six months (vs Google 418M and Meta 227M for the year) | 3 billion · 300,000+ derivatives | Hugging Face, 14 Aug 2026 |
| Cheapest GPT-4-class inference | Down ~60x in 45 months | State of Open Source AI v1.1 |
A 7–9 point gap on the overall index sounds small, but the 92-Elo gap on expert knowledge work tells the other half. On work that needs professional judgment, the top closed models still lead clearly. On well-scoped routine work, Chinese open models get the job done at far lower cost. That is why usage keeps flowing to China without the test scores having to win.
"Open" no longer means free: Kimi K3 ships under Moonshot's own license. Any provider earning more than $20 million a year from the model needs a separate agreement, and reporting says Moonshot can ask for up to 30% revenue share. Reuters reported on 7 August that Alibaba will apply the same approach to the next Qwen release, unlike Qwen3's Apache 2.0. DeepSeek moved to time-of-day pricing on 16 August, which the State of Open Source AI report counts as the first list-price rise by an open-weight lab, and Zhipu itself says it raised API prices about 101% in the first half. Any organization choosing Chinese models because they are "free" needs to read the next release's license first. The same report notes that of 16 notable open releases it examined, none published its training-data recipe under the open-source definition.
3. Anthropic's 10 September Report: Seven Chinese Labs, 190 Million Exchanges, and Beijing Summoning All Seven
The story that hit Chinese models' credibility hardest this month came from Anthropic's fourth threat-intelligence report, published on 10 September 2026. It states that between December 2025 and August 2026, seven Chinese companies used fraudulent accounts and proxy networks across several countries to call Claude roughly 190 million times and use the answers to train their own models, which Anthropic calls illicit distillation.
| Company | Exchanges cited by Anthropic | Window | Additional detail in the report |
|---|---|---|---|
| Alibaba | More than 151 million | May–Jul 2026 | The largest volume in the report |
| Moonshot (Kimi) | About 23 million | May–Jul 2026 | In one roughly 10-day window, nearly 300,000 requests via 5,380 fraudulent accounts, mostly appearing from Singapore and Japan · Anthropic says Moonshot "silently forwarded" Kimi users' requests to Claude "and showed users the answers as though they were Kimi's" · one user Anthropic assessed as affiliated with the Chinese military asked it to analyze footage from hundreds of CCTV cameras in Chengdu |
| DeepSeek | More than 12.1 million | 14 days in Jul 2026 | Fed conversations between its own model and its users into Claude |
| Zhipu | 3.4 million | — | |
| Xiaomi | More than 400,000 | — | |
| MiniMax, SenseTime | No figure given | — | Named in the report |
Source: Anthropic, Countering misuse of AI: September 2026 (10 Sep 2026), as summarized by The Next Web, The Hacker News and Quartz · all counts are Anthropic's own
The technique the report dwells on is "cross-session chain-of-thought replay": taking the thinking blocks Claude returns and replaying them across conversations to pull out its internal reasoning. That is the direct reason Claude models from Opus 5.5 and Sonnet 5.5 onward bind thinking blocks to the account and conversation that produced them, and added a reasoning_extraction refusal category. The two stories are one story.
What followed matters more than the report itself. On 22 September, the Cyberspace Administration of China (CAC) summoned all seven companies, and reporting says the investigation focused on DeepSeek and Moonshot. The reason was not Anthropic's intellectual property but that the examples in the report suggested sensitive Chinese security data had been sent to foreign servers. As of this writing there are no penalties, and neither Moonshot nor DeepSeek has issued a statement.
What Thai organizations should take from this: the point is not whether Chinese models are good or bad. It is whether you know whose servers your staff's prompts end up on. In the Kimi case as reported, users thought they were talking to a Chinese model while their data went to the United States through accounts in Singapore. The reverse also happens: an app that claims a Western model may route through a provider you have never heard of. The only protections are a contract that names the data processor, or running open weights on your own hardware. We cover the second option in Running DeepSeek in-house and the risk picture in DeepSeek and China AI risks.
4. Alibaba's Apsara Conference, 22 September: Qwen 4, In-House Chips, and a 20-Gigawatt Target
The Apsara Conference in Hangzhou is the event Alibaba plans its whole year around. This year the substance came in three layers: models, chips, and data centers.
- Models: Qwen 4 is in training on a new-generation architecture and was previewed in four tiers (Max, Plus, Flash, 27B) with no release date · Qwen 4.5 and Qwen 5 target 5–10 trillion parameters · Alibaba says Qwen3.8-Max went through 33 rounds of Recursive Self-Improvement, lifting its Artificial Analysis score from 40 to 45 · a live-translation model and an audio model shipped alongside
- Chips: Zhenwu V900 with 216 GB of memory and 1,200 GB/s inter-chip bandwidth, three times the performance of the M890, mass production in Q1 2027 · Yitian 720 and 730 CPUs in 2027 · Alibaba says Zhenwu chips already serve more than 650 customers
- Data centers: a target of more than 20 gigawatts of global Alibaba Cloud capacity by 2032 · the next day it announced new regions in Türkiye, Finland and the Netherlands and expansions in Malaysia, Germany, the UAE, France and Hong Kong, for 107 zones across 31 regions
- Consumer hardware: Qwen Book, a computer whose operating system is driven by an agent; Qwen Glasses N1; Qwen Clip earbuds co-engineered with Bose; and the QwenNote A2 recorder at RMB 1,199, China only
CEO Eddie Wu's line that "machine thinking still has an enormous growth runway," set against the 20-gigawatt figure, says Alibaba is not competing on models alone but on owning the whole stack from chip to app. For Thai organizations the tangible piece is that Alibaba Cloud held Qwen Conference Thailand in Bangkok on 4 September, with nearly 400 attendees from insurance, finance, consulting and media. That is direct enterprise selling in Thailand, no longer just outreach to developers.
5. Chips: Huawei Pulls Its Roadmap Forward Three Quarters, and Nvidia's China Share Falls Below 10%
This month's chip news has to be read as a pair. The United States has allowed case-by-case sales of the H200 to China since 15 January 2026, with a 25% levy on revenue. But Chinese authorities told customs not to let the H200 into the country from mid-January and called technology companies in to say "do not buy unless necessary." The result is a Chinese AI-chip market walled off for domestic suppliers by both governments at once.
| Item | Figure / schedule | Source |
|---|---|---|
| Ascend 960DT (training) | Q1 2027, three quarters ahead of plan · up to 288 GB memory · about 4 petaFLOPS at FP4 | Huawei Connect 2026, TrendForce 17 Sep |
| Ascend 960PR (inference) / 970 / 980 | Q3 2027 / 2028 / 2029, one generation a year | Huawei Connect 2026 |
| Atlas 960E SuperPoD | 4,096 NPUs · 8 EFLOPS at FP8 · 1 petabyte of HBM per pod (down from a planned 15,488 chips) | The Next Web 18 Sep |
| China AI-chip market share, 2026 forecast | Huawei ~50% (~$12.1B) · Nvidia ~8% (~$2.0B), down from ~40% in 2025 | Bernstein |
| Actual output | Huawei will produce under 4% of Nvidia's AI compute in 2026 · Huawei itself expects to meet Chinese demand around 2030 | Epoch AI · Huawei Connect |
| Model-lab customers | DeepSeek plans to deploy more than 160,000 Ascend 950DT chips · ByteDance, Alibaba and Tencent have placed large Ascend orders | The Next Web · Bernstein |
| Developers | External developers are 61% of the CANN community, outnumbering Huawei staff for the first time · Kunpeng ecosystem 4.16 million developers, 7,200 partners | Huawei Connect |
The right reading is that Huawei wins on "share of the Chinese market" because its competitor is locked out, but still loses on "total output" by a wide margin. That means the next generation of Chinese models will be trained and served on scarce chips. The August price increases at DeepSeek and Zhipu are not a coincidence. They are the signal that cheap Chinese models may not stay cheap.
6. The Money: Zhipu Quadruples Revenue but Still Loses RMB 2 Billion; Moonshot Valued at $35 Billion
Zhipu (Z.ai internationally) is the only listed Chinese model company, so it is the one window into the real economics of the business. Its H1 2026 results, reported on 31 August, look like this.
| Item | H1 2026 | Note |
|---|---|---|
| Total revenue | RMB 957 million (~$142 million), up nearly 400% | |
| API / open-platform revenue | RMB 825 million, up 27x, 86.5% of revenue | 15.2% a year earlier; the business flipped from projects to selling tokens |
| Net loss | RMB 2,072 million (down 12.5%) | API gross margin 24.6%, from negative a year earlier |
| ARR at end of August | $1.6 billion (monthly × 12) · above $2 billion on a weekly basis after GLM-5.3 | 7.4 million MaaS users, up 144% · 115 customers above $100,000 ARR |
| R&D spend | RMB 2,130 million | More than the half-year's revenue |
| September raise | ~$5 billion via new shares and zero-coupon convertibles due 2027 · ~60% for the next-generation model | After a $4 billion placement in July |
| Next generation | GLM-6.0, a "Full Self-Training" approach, no date | Founder Tang Jie, on the investor call |
Sources: Zhipu earnings-call minutes as reported by 36Kr (2 Sep 2026) and the September 2026 placement announcements
These numbers say two things at once. Token revenue is growing fast and now has a positive gross margin, but R&D still exceeds revenue and the company raises billions of dollars every couple of months. Moonshot closed its $3.5 billion round at a $35 billion valuation in July after Kimi K3 shipped, then two months later was accused by Anthropic and summoned by the CAC. The risk with Chinese vendors is not technical. It is business continuity and regulation.
7. Rules: China Pulls 14,000 AI Apps and Regulates Emotional Chatbots
China regulates AI through a series of measures rather than a single law. Two items took effect in this period.
- The 2026 Qinglang campaign — the CAC began it in April. First-phase results published on 6 July: more than 14,000 AI products removed for skipping model registration or safety review, more than 6 million illegal posts deleted, 26,000 accounts suspended, and 9 open datasets taken down. The second phase, with results published on 2 September, extended to seven content categories including fake news impersonating state media, deepfakes of public figures, AI "resurrection" of the deceased, and unregistered shell apps.
- The Interim Measures on Anthropomorphic AI Interactive Services — issued 10 April, effective 15 July 2026. They cover services offering "continuous emotional interaction simulating a real person's personality," requiring a minors mode, usage time limits, real-world reminders, and guardian control over spending. Q&A systems, work assistants and customer service are excluded.
For Thai users, the direct effect is that Chinese AI apps serving China must register their models and label AI-generated content, a burden the providers already carry. Services offered outside China are not covered. A Thai organization using Chinese models through a cloud outside China, such as Alibaba Cloud in Singapore or Malaysia, is governed by its contract and Thai law instead. For the European side, see our piece on the EU AI Act and Thai exporters.
So What Should Thai Organizations Do with This Month's News?
Four things you can do in October
1. Inventory which model every AI app your staff use actually connects to — the lesson of the Kimi case is that the app's name does not tell you where the data goes. Ask the vendor directly who the downstream data processor is and in which country, and write it into the contract.
2. If you use Chinese models, pick one of two clear paths — run open weights on your own hardware (data stays in, but check the new revenue-sharing licenses), or use them through a cloud with a proper contract. Do not use consumer apps for work involving customer data.
3. Cost with peak-hour prices, not the headline price — DeepSeek charges full rate during Thai working hours, so daytime work is less of a bargain versus Western models than the list suggests, and both DeepSeek and Zhipu have already raised prices in August.
4. Do not lock your systems to one vendor — a Chinese vendor can be summoned by its regulator within two weeks; a Western vendor can block access from a given country. Your integration layer must let you change models with one setting.
Conclusion
- Chinese models this month: DeepSeek V4.1-Flash on a new architecture at lower prices · Xiaomi MiMo-V2.6-Pro scored 46, the top open-weight model · Qwen 4 in training, Qwen 5 targeting 10 trillion parameters.
- Chinese open models trail the top closed models by 7–9 points on the overall index but hold 7 of the top 10 slots by usage on OpenRouter.
- Anthropic accused seven Chinese labs of harvesting 190 million exchanges, Alibaba the largest at 151 million · the CAC summoned all seven on 22 September · no response from the accused yet.
- "Open" is not free: Kimi K3 takes a revenue share above $20 million, the next Qwen will follow, and DeepSeek and Zhipu have raised prices.
- Huawei pulled the Ascend 960DT to Q1 2027 and is forecast at 50% of China's market, but its output is under 4% of Nvidia's. Scarce chips mean Chinese model prices may not stay low.
- Zhipu quadrupled H1 revenue with API at 86.5%, still lost RMB 2 billion, and raised another $5 billion.
- Thai organizations: map the data path of every AI app, cost with peak prices, and design the integration layer so the model can be swapped.
This month Chinese models did not win on scores. They won on volume and lost on trust. For an organization the question is not whose model you use, but where your data ends up.
- Paitoon Butri · Network & Server Security Specialist, Grand Linux Solution Co., Ltd.
References
- Anthropic — Countering misuse of AI: September 2026 (10 Sep 2026)
- The Next Web — Beijing turns on DeepSeek and Moonshot over Claude data (22 Sep 2026)
- The Hacker News — Anthropic Says Seven China-Based AI Labs Ran Industrial-Scale Claude Distillation (Sep 2026)
- DeepSeek API Docs — Change Log
- DeepSeek API Docs — Models & Pricing
- VIR — Alibaba targets 10 trillion parameters with next-generation Qwen 4 model (Sep 2026)
- TechNode Global — Alibaba unveils Qwen Book agentic computer and AI wearables at Apsara 2026 (24 Sep 2026)
- Alibaba Cloud — Expands Global Infrastructure and AI Portfolio to Accelerate Enterprise AI Adoption (23 Sep 2026)
- Alibaba Cloud Community — QwenCloud at Qwen Conference Thailand 2026
- Fortune — Alibaba AI models hit 3 billion downloads, passing Meta, Google (15 Aug 2026)
- AI News — Alibaba tests new business model for Qwen open-source AI (7 Aug 2026)
- Tom's Hardware — Alibaba claims new Qwen Image 2.1 beats Google Nano Banana 2.0 (Sep 2026)
- TechNode — Xiaomi open-sources MiMo-V2.6 models after scaling reinforcement learning (22 Sep 2026)
- Fortune — Moonshot's Kimi K3 pushes Chinese AI into Fable-level territory (16 Jul 2026)
- Bloomberg — Moonshot AI Surpasses Funding Goal to Hit $35 Billion Value (29 Jul 2026)
- 36Kr — Zhipu Earnings Call Minutes: Revenue Structure Reversed (2 Sep 2026)
- The Next Web — Huawei Connect 2026: Ascend 960 early, a million-NPU plan (18 Sep 2026)
- TrendForce — Huawei Speeds Up AI Chip Roadmap, Reportedly Pulls Ascend 960DT Forward Three Quarters to 1Q27 (17 Sep 2026)
- MarketScale — Nvidia's China AI chip share is forecast to collapse from 40% to 8% as Huawei scales (Bernstein)
- Epoch AI — Will Huawei catch up to Nvidia by 2030?
- Introl — BIS Export Policy Shift: H200 and MI325X case-by-case (Feb 2026)
- Tom's Hardware — Chinese customs told to block H200 imports (Jan 2026)
- State of Open Source AI v1.1 (September 2026)
- Yahoo News — China Purges Over 14,000 AI Products Amid Qinglang Cleanup (6 Jul 2026)
- Bird & Bird — China's New Regulations on AI Anthropomorphic Interactive Services (2026)
Want AI on your business data, and to know where that data goes?
The Grand Linux team designs the AI-to-ERP integration layer over MCP: permissions by role, a log of every call, and a model endpoint you can swap without rebuilding, whether it is a Western model or open weights running on your own hardware.
Request a Free DemoTel 02-347-7730 | sale@grandlinux.com
