DeepSeek V4–almost on the frontier, a fraction of the price
Chinese AI company DeepSeek has released two preview models of its DeepSeek V4 series, DeepSeek-V4-Pro and DeepSeek-V4-Flash, both featuring a 1 million token context length and built using a Mixture of Experts architecture. The models are notable for their competitive performance against frontier AI systems while being significantly cheaper to operate, with DeepSeek-V4-Flash priced lower than comparable small models from OpenAI, Google, and Anthropic. DeepSeek attributes the cost efficiency to substantial improvements in computational efficiency, particularly for long-context tasks, as demonstrated in internal benchmarks and technical metrics.
Opening excerpt (first ~120 words) tap to expand
DeepSeek V4—almost on the frontier, a fraction of the price 24th April 2026 Chinese AI lab DeepSeek’s last model release was V3.2 (and V3.2 Speciale) last December. They just dropped the first of their hotly anticipated V4 series in the shape of two preview models, DeepSeek-V4-Pro and DeepSeek-V4-Flash. Both models are 1 million token context Mixture of Experts. Pro is 1.6T total parameters, 49B active. Flash is 284B total, 13B active. They’re using the standard MIT license. I think this makes DeepSeek-V4-Pro the new largest open weights model. It’s larger than Kimi K2.6 (1.1T) and GLM-5.1 (754B) and more than twice the size of DeepSeek V3.2 (685B). Pro is 865GB on Hugging Face, Flash is 160GB. I’m hoping that a lightly quantized Flash will run on my 128GB M5 MacBook Pro.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at Simon Willison’s Weblog.
Discussion
0 commentsMore from Simon Willison’s Weblog
