3 Minutes
DeepSeek has pulled the curtain back on DeepSeek-V4 Preview, and the numbers are hard to ignore. The Chinese AI company is now offering two new models, V4 Pro and V4 Flash, both built for one million token context windows, a benchmark that puts long-document handling, code analysis, and complex reasoning squarely in the spotlight.
The models are already available through DeepSeek’s website in Instant Mode and Expert Mode, while the API has also been updated and is live today. In a market where context length has become one of the fiercest battlegrounds in AI, DeepSeek is making a very direct play: offer serious capability without the premium price tag.
Two models, one very big promise
On paper, the difference between the two variants is striking. DeepSeek-V4 Pro carries 1.6 trillion total parameters with 49 billion active parameters, while V4 Flash comes in leaner at 284 billion total parameters and 13 billion active parameters. Both support the same massive 1M context length, but they are clearly aimed at different users and budgets.
DeepSeek says the Pro model is the heavyweight in the family. It is designed with stronger agentic abilities, broader world knowledge, and advanced reasoning that the company claims outperforms current open models across math, STEM, and coding. DeepSeek also says Pro is competitive with top closed-source systems, though it places a note of caution by saying it still trails Gemini 3.1 Pro in some areas.
Flash, meanwhile, is the sharper value play. DeepSeek describes it as delivering reasoning that comes close to Pro, while matching Pro on simpler agent tasks. The real appeal is cost. Flash is positioned as the more affordable option for developers who want long-context AI without burning through their budget.
The pricing reflects that strategy. For Flash, input costs start at $0.028 with a cache hit and $0.14 with a cache miss, while output is priced at $0.28. Pro sits much higher, with input priced at $0.145 or $1.74 depending on cache status, and output priced at $3.48.
For users who want to try the models right away, DeepSeek has opened access at chat.deepseek.com. The company also says the open weights are available, along with a technical report for anyone who wants to dig into the architecture, benchmarks, and training details.
It is a bold release, and one that pushes the long-context conversation forward again. DeepSeek is not just chasing headline numbers. It is trying to prove that massive AI models can still be practical, accessible, and affordable.
Source: huggingface
Leave a Comment