DeepSeek V4 Preview: Two Sizes, One Aggressive Price Sheet
DeepSeek previewed two new models this week: V4 Flash and V4 Pro. Both are mixture-of-experts architectures with 1 million token context windows. Both undercut major Western competitors on price. Neither supports images, audio, or video.
The Parameters
V4 Pro is large. 1.6 trillion total parameters, 49 billion active. DeepSeek describes it as the biggest open-weight model currently available, ahead of Moonshot AI's Kimi K 2.6 (1.1 trillion parameters) and MiniMax M1 (456 billion parameters). The previous flagship, V3.2, had 671 billion parameters.
V4 Flash runs leaner: 284 billion total parameters, 13 billion active.
Benchmarks
V4 Pro Max outperforms open-source peers on reasoning benchmarks. It beats GPT-5.2 and Gemini 3.0 Pro on some tasks. Coding competition benchmarks put both V4 models roughly on par with GPT-5.4.
The weaker area is knowledge tests. Both V4 models trail GPT-5.4 and Gemini 3.1 Pro there by an estimated 3 to 6 months of developmental trajectory. That gap is real.
Pricing
V4 Flash: $0.14 per million input tokens, $0.28 per million output tokens. That undercuts GPT-5.4 Nano, Gemini 3.1 Flash, GPT-5.4 Mini, and Claude Haiku 4.5.
V4 Pro: $0.145 per million input tokens, $3.48 per million output tokens. That undercuts Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.7, and GPT-5.4.
1.6 trillion parameters for less than Gemini 3.1 Pro's price. Whatever else is happening here, the unit economics are not subtle.
The Caveats
These are previews. Benchmark performance and production performance are different things.
Text only. Teams that need multimodal inputs will need to look elsewhere for now.
The Pattern Continues
DeepSeek has done this before: ship a large open-weight model priced to make Western frontier labs uncomfortable. V4 follows V3.2 in that tradition, with substantially more parameters and similar pricing aggression.
Closing the coding benchmark gap with GPT-5.4 while undercutting it on price is a meaningful data point. The 3 to 6 month knowledge lag is a real limitation. Whether that matters depends entirely on the use case.
Wait for the production release before drawing strong conclusions.
Source: Techcrunch