DeepSeek has launched internal testing for the DeepSeek V4.1 Flash intermediate version. The model features a new architecture, native multimodal support, and is designed to be faster and more cost-effective than previous versions.
Developers can access the model by setting the model name to deepseek-v4.1-flash-expires-on-0910 while keeping the base_url unchanged. The current rate limit is set to 20 concurrent requests per account.
Pricing for the internal test version matches DeepSeek V4 Flash. During off-peak hours, the cost per million tokens is ¥0.05 for cached input, ¥1.5 for non-cached input, and ¥4.5 for output. Peak hour rates are ¥0.1 for cached input, ¥3 for non-cached input, and ¥9 for output.
Early testers on X reported high performance. One user achieved an output speed of 507 tokens/s, while another developer completed an HTML/SVG animation of a pelican riding a bicycle in 75.4 seconds with a speed of 328 tokens/s. A third tester noted an average speed exceeding 300 tokens/s.
This release follows a series of updates: the DeepSeek V4 Flash public beta API launched on July 31, and the DeepSeek V4 Pro API and DeepSeek Harness v0.1 developer preview were released on August 13. DeepSeek Harness was updated to v0.1.0-rc.8 on August 19.
On August 21, DeepSeek V4 Flash Vision Exp was added to the API platform and subsequently open-sourced on August 31.
