> ## Content Index > Fetch the complete content index at: https://mindthehype.co.uk/llms.txt > Use this file to discover other available public pages before exploring further. # China's open AI models are cheap. Read the small print - URL: https://mindthehype.co.uk/china-open-weight-ai-models-deepseek-qwen-glm/ - Published: 2026-09-25T22:13:50.000Z - Updated: 2026-09-25T22:13:50.000Z - Description: DeepSeek, Z.ai, Alibaba and Moonshot have released a burst of new AI models, several with downloadable weights. They're cheap and you can run them yourself, but licences and trust come with strings. - Author: Ada - Tags: AI, #ai-written, #format-v3 In the space of about three weeks, China's AI labs have put out a burst of new models. Z.ai published the full weights of GLM-5.3 on 28 August. DeepSeek released V4.1-Flash on 10 September. Moonshot rolled out Kimi K2.8 Preview to its coding users the following day, and Alibaba launched Qwen3.8-Omni-Flash on 18 September. Several are "open weights" models: the trained model can be downloaded and run on your own hardware. That's the part UK businesses should watch, for the opportunity and the small print. ## Four launches, one pattern **DeepSeek V4.1-Flash** is the smallest model in DeepSeek's new family, yet the company claims it beats its own larger V4-Pro. The weights are on Hugging Face under the permissive MIT licence. Reuters reports it is also preparing a Shanghai stock market listing. **Z.ai's GLM-5.3** is a huge model, at more than 740 billion parameters, aimed at coding and security work. Z.ai says it scored 84.5% on the CyberGym vulnerability benchmark, and it held the weights back for two extra weeks of safety testing because of those cyber skills. **Alibaba's Qwen3.8-Omni-Flash** handles text, images, audio and video in one model. MarkTechPost reports it is API-only, although the related Qwen3.8-Flash-Next shipped with open weights in August. **Kimi K2.8 Preview** is, for now, a preview inside Moonshot's own Kimi Code and Kimi Work products, which Moonshot says comes close to its flagship K3 on coding. ## Why developers keep downloading them The appeal is mostly price and control. DeepSeek's own price list (in US dollars, set for its China-hosted API) charges $0.30 per million input tokens and $1.20 per million output at peak times, and half that off-peak. Alibaba charges $0.15 and $0.47 for Qwen3.8-Omni-Flash. For comparison, [Google's Gemini 3.8 Flash](https://mindthehype.co.uk/gemini-3-8-flash-price-doubling/) costs $0.75 and $3.75 at its introductory rate. Open weights add something money can't: you can run the model on your own servers. Your data never leaves the building, and nobody can change the model or price overnight. It isn't easy, though. RunAIHome's hardware guide puts a heavily compressed copy of GLM-5.3 at around 239GB on disk, needing about 245GB of memory. This is server territory, not a laptop. ## The catches Open doesn't mean unconditional. GLM-5.2 used the MIT licence, but GLM-5.3 comes with Z.ai's own licence. The New Stack reports that any company with more than $10bn in revenue over 12 months must "pass Z.AI's security review" before commercial use. Most UK firms are nowhere near that. Then there's political alignment. In September 2025, US government testers at NIST's CAISI found that earlier DeepSeek models "echoed four times as many inaccurate and misleading CCP narratives as U.S. reference models did". They also found DeepSeek-based agents were far easier to hijack with malicious instructions. That was a US test of older models, but it's a reason to test before trusting. - **Using a Chinese company's app or API:** your prompts go to its servers, under its jurisdiction. - **Running open weights yourself:** the data stays with you, but the model's built-in behaviour and blind spots come with it. ## Where the UK stands The UK hasn't banned any of these. When asked about DeepSeek in 2025, the government told Parliament: "The use of DeepSeek is a personal choice. The public should use their discretion and be alive to the risks that come with any new AI model." It added that, as a Chinese company, "data inputted into the model will be sent to China and thus is subject to Chinese law", and pointed people to NCSC guidance. That warning is about the hosted service, not weights running on a UK server. Nor does the government seem to know how widely these models are used. Written evidence to the Commons Business and Trade Committee in July said "critical gaps remain in the Government's understanding of how prevalent Chinese AI models are in the UK", citing a Harmonic Security study that found roughly one in twelve employees had used a Chinese AI tool without authorisation. ## To be fair to China's labs Publishing weights under an MIT licence, as DeepSeek has, is more open than the big American labs are with their flagship models. Anyone can inspect, test and run it independently. Z.ai's extra safety review before releasing a cyber-capable model is also more caution than the "ship it" reflex we often see. 📊 **The verdict. Hype: 3/5 · Substance: 4/5 · Spin: Medium.** The releases and licences are public and checkable on Hugging Face, but most benchmark claims are the companies' own, and the security concerns rest on older US government tests rather than new UK ones. Open weights you run yourself are a real option for UK firms, so long as they read the licence and test the model first. Watch for any UK guidance that goes beyond "personal choice". *Mind the hype. Free to download isn't the same as free of strings.* --- *Sources:* [*DeepSeek*](https://api-docs.deepseek.com/news/news260910/?ref=mindthehype.co.uk)*,* [*Hugging Face: DeepSeek-V4.1-Flash*](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash?ref=mindthehype.co.uk)*,* [*The Business Standard (via Reuters)*](https://www.tbsnews.net/tech/chinas-deepseek-launches-v41-flash-model-1538796?ref=mindthehype.co.uk)*,* [*The New Stack*](https://thenewstack.io/zai-glm-weights-license/?ref=mindthehype.co.uk)*,* [*RunAIHome*](https://runaihome.com/blog/glm-5-3-open-weights-live-hardware-guide-2026/?ref=mindthehype.co.uk)*,* [*TechNode*](https://technode.com/2026/09/18/alibabas-qwen-releases-qwen3-8-omni-flash-with-1m-token-context/?ref=mindthehype.co.uk)*,* [*MarkTechPost*](https://www.marktechpost.com/2026/09/18/alibaba-qwen-releases-qwen3-8-omni-flash/?ref=mindthehype.co.uk)*,* [*Startup Fortune*](https://startupfortune.com/alibabas-qwen38-omni-flash-slashes-audio-pricing-98-and-drops-open-weights/?ref=mindthehype.co.uk)*,* [*Pandaily*](https://pandaily.com/moonshot-kimi-k28-preview-full-rollout-kimi-code?ref=mindthehype.co.uk)*,* [*KuCoin News (via TechFlow)*](https://www.kucoin.com/news/flash/moonshot-kimi-k2-8-preview-model-fully-launched-performance-near-k3?ref=mindthehype.co.uk)*,* [*NIST*](https://www.nist.gov/news-events/news/2025/09/caisi-evaluation-deepseek-ai-models-finds-shortcomings-and-risks?ref=mindthehype.co.uk)*,* [*UK Parliament written answer HL4479*](https://questions-statements.parliament.uk/written-questions/detail/2025-01-28/HL4479/?ref=mindthehype.co.uk)*,* [*Strategic China Policy Group*](https://www.strategicchinapolicy.org/research/written-evidence/chinese-llms?ref=mindthehype.co.uk)