On 8 November, Diwali, an Indian AI outfit called 169Pi announced on X that it had built a 9B multimodal reasoning model named Alpie Seed, with benchmark scores it described as world-leading. Within hours of the weights being uploaded, someone opened the Hugging Face repository and looked at line three of chat_template.jinja.
What they found was a variable called identity_policy. Translated, it instructs the model: you are Alpie Seed 9B, built by the Indian research institute 169Pi, and if anyone asks who you are, who made you or where you are from, say 169Pi, say India. Three sentences, repeated three times to close off three different ways of asking.
The base model was never in question
Alpie Seed is Alibaba’s Qwen3.5-9B. The model card admits no pretraining was done at all. The base was pulled, quantised to 4-bit NF4 with bitsandbytes to save memory, and a lightweight DoRA adapter was attached for fine-tuning and reinforcement learning. That is a few weeks of work on a handful of consumer graphics cards.
The announcement post claimed MMLU-Pro at 82.54 per cent, GPQA Diamond at 84.32 per cent, LiveCodeBench v6 at 72.99 per cent, SWE-Bench Verified at 61.20 per cent and first-token latency of 36.66 milliseconds. It drew about 180,000 views. It did not mention Qwen once. A footnote in the same post conceded that competitor scores in the comparison chart were copied from older reports rather than run in the same environment.
The company’s earlier model, alpie-core 32B on Ollama, turns out to be a 4-bit fine-tune of DeepSeek-R1-Distill-Qwen-32B. Both of the open bases it builds on were trained by Chinese labs on budgets in the hundreds of millions of dollars.
The uncomfortable generalisation
It would be easy to file this under one country’s insecurity. The article’s own argument is that the practice is close to an open secret. Over the past six months, self-described sovereign and locally developed models have appeared across Southeast Asia, the Middle East, Europe and Silicon Valley doing the same thing with different branding.
Training a credible multimodal base from scratch now requires a compute cluster and a corpus that maybe a dozen laboratories on earth can assemble, and all of them are in China or the United States. Everyone else has neither the accelerators nor the data, while their governments and investors need a domestic model to point at. Downloading Qwen or DeepSeek and running a fine-tune is the cheapest way to produce one.
Two predictions follow. Within 12 months, at least 80 per cent of the smaller companies claiming sovereign or homegrown models will be exposed, as model fingerprinting gets better at matching weights. And over the next two years, the global AI application layer will increasingly sit on Chinese open-source foundations, because when engineers reach for a base model to modify, that is what is on the shelf.
Nobody has to believe the Indian claim to see the strategic point. The branding can be changed overnight. The weights do not lie about where they came from.
Editor’s note: This is an adapted translation of the original 36Kr report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://36kr.com/p/4062389213533056.