DeepSeek’s removal of R1 references from its chatbot is fueling speculation over the fate of its anticipated R2 AI model.

The Chinese AI startup recently released an updated version of its V3 model, V3.1, expanding the context window to 128,000 tokens. This allows the model to process around 300 pages of text in a single interaction.

The update was shared with a WeChat user group rather than through public channels.

The startup has not provided details about future model development, leaving the status of the R2 model unclear.

Independent tests show that V3.1 has improved in coding tasks, but it still lags behind leading global models.

Some users report little progress in reasoning abilities and mixed results in text generation quality.

🔗 Source: South China Morning Post

🧠 Food for thought

1️⃣ Hardware sovereignty pressures create strategic compromises for Chinese AI companies

DeepSeek’s R2 model delays reveal how government technology mandates can clash with technical realities in AI development.

The company was urged by Chinese authorities to train its next-generation model using Huawei’s Ascend chips, but encountered persistent technical failures that forced them to revert to Nvidia hardware for training while keeping Huawei chips for inference tasks 2.

This compromise highlights a broader challenge facing Chinese AI firms: balancing domestic technology requirements with the need to maintain competitive performance in global markets.

The situation reflects the technical gaps that still exist in China’s semiconductor ecosystem, where domestic alternatives haven’t yet matched the reliability and performance of established players like Nvidia, particularly for demanding AI training workloads.

2️⃣ Early cost advantages require sustained innovation momentum to maintain market position

DeepSeek’s declining market share demonstrates how initial efficiency breakthroughs can quickly lose impact without consistent follow-through.

The company originally gained attention by training its R1 model for just $6 million compared to competitors spending significantly more, but their share on Chinese cloud platform PPIO dropped from 99% to 80% as Alibaba’s Qwen models gained traction 1.

Recent DeepSeek developments