Another Contender Enters the Ring
DeepSeek has officially dropped V4, its follow-up to the model that shook Silicon Valley in early 2025. The Chinese AI lab released two variants: DeepSeek-V4-Pro (1.6T parameters, 49B activated) and DeepSeek-V4-Flash (284B parameters, 13B activated). Early benchmarks suggest the Pro variant trades blows with Claude Sonnet 4.5 and approaches Opus 4.5 territory on reasoning and coding tasks.
What Changed Under the Hood
The V4 series introduces a refined Mixture-of-Experts architecture with improved routing efficiency. DeepSeek claims a 40% reduction in inference cost compared to V3 while delivering better accuracy across MATH, HumanEval, and MMLU-Pro benchmarks. The Flash variant targets the same performance tier as GPT-5.5-mini but at roughly a third of the compute budget.
- Pro: 1.6T total — 49B active per token
- Flash: 284B total — 13B active per token
- Context window: 256K tokens (both variants)
- License: MIT (open-weight, Apache 2.0 for commercial use)
The Geopolitical Angle
DeepSeek's timing is notable — V4 landed hours after OpenAI's GPT-5.5, a deliberate escalation in what the CFR calls "a new phase in the U.S.-China AI rivalry." Anthropic and OpenAI have both alleged that DeepSeek engaged in industrial-scale distillation, creating over 24,000 fake accounts to extract training signals from Claude models. Whether V4's performance reflects genuine architectural advances or aggressive distillation is the subject of heated debate.
Regardless, the practical implication is clear: open-weight models now compete with closed frontier systems on price, if not always on polish. For teams self-hosting on GPU clusters, V4-Flash at $0.15/M tokens offers a compelling alternative to API-based providers.
Comments