DeepSeek V4 AI: Crushing The Competition

TL;DR

DeepSeek V4 matches frontier models at a fraction of the cost using three-layer KV-cache compression that cuts memory needs by 90%.

Key Points

• 1.DeepSeek V4 Pro matches billion-dollar frontier models while being open and free. It supports a 1 million token context window — roughly 1,500 pages of dense documentation — a feature previously exclusive to Google's Gemini flagship.
• 2.Three compression layers slash KV-cache memory by ~90%. Token-level compression summarizes text, Heavily Compressed Attention applies 128-to-1 structural compression, and Compressed Sparse Attention acts like an index — together reducing storage needs without major information loss.
• 3.The new Pro model uses 3x less compute than its predecessor; the Flash model uses 10x less. Despite heavy compression, recall benchmarks show the Pro version outperforms Google's Gemini 3.1 Pro on retrieving hidden facts in long contexts.
• 4.Pricing is 8–30x cheaper than Anthropic's Claude depending on discounts. The model can be self-hosted or accessed online, and the host notes that 'intelligence will soon get too cheap to meter.'
• 5.Key limitations include unimodal input (no images or audio), context-window degradation, and unexplained training stabilization techniques. The creators themselves admit they don't fully understand two techniques that stabilize training, though the paper is praised for its transparency.