DeepSeek V4
DeepSeek launched V4 as a Preview in two open-weight Mixture-of-Experts variants — V4-Pro (1.6T total / 49B active params) and V4-Flash (284B total / 13B active) — positioned as the successor to V3 and R1. The company claimed V4-Pro delivers performance rivaling top closed-source models, leading open models in world knowledge (trailing only Gemini 3.1 Pro) and in math, STEM, and coding. DeepSeek emphasized a 1M-token context as the new default and a novel sparse-attention design for efficiency, alongside hybrid Thinking / Non-Thinking modes in a single model. As with prior generations, the pitch was frontier-adjacent capability at a fraction of incumbent pricing.
V4 shipped as promised: open weights under a permissive (MIT-style) license on Hugging Face, plus immediate API access compatible with both OpenAI and Anthropic interfaces, with 1M context as the default. Pricing came in aggressively low — roughly $0.14/$0.28 per million input/output tokens for Flash and about $1.74/$3.48 for Pro — and legacy deepseek-chat/deepseek-reasoner endpoints were slated for retirement on 2026-07-24.
DeepSeek again did the thing that unsettles incumbents: it shipped near-frontier capability with open weights at a fraction of closed-model pricing, and it actually delivered the weights and API on day one. The honest caveat is the 'Preview' label — this is an early release, and the headline 'rivals top closed models' claim holds best on knowledge and STEM benchmarks rather than uniformly across every task. The bigger picture is that the open-weight frontier is now substantially Chinese-led, which carries real procurement, data-governance, and geopolitical considerations for Western enterprises even when the license is permissive. Cost-wise the disruption is genuine: at these token prices the economics of high-volume AI workloads shift materially. Executives should treat V4 as proof that 'good enough, open, and cheap' is a durable competitive vector, while weighing where it is and isn't appropriate to deploy.
V4 keeps compressing the price of near-frontier AI and reinforces that the leading open-weight models increasingly come from China, forcing buyers to weigh dramatic cost savings against governance and sourcing questions.