DeepSeek V4 Preview: Huge Specs, Tiny Price, Mixed Real-World Results
Wait, the cheap giant? DeepSeek V4 Preview lands with Pro and Flash, and the pitch is huge: open weights, MIT license, and a Pro model claiming a 1 million token memory window. On specs, it sounds absurd.
Wait, the cheap giant? DeepSeek V4 Preview lands with Pro and Flash, and the pitch is huge: open weights, MIT license, and a Pro model claiming a 1 million token memory window.
On specs, it sounds absurd. Pro lists 1.6 trillion total parameters with 49 billion active. Flash is smaller at 284 billion total and 13 billion active, yet built to be faster and far cheaper.
Then reality hits. In browser tests, coding and interface work often looked rough, unfinished, or just off. A desktop clone felt flimsy. A controller looked wrong. Sometimes the cheaper Flash even felt better than Pro.
That is the twist. Pro costs 14 cents per million input tokens and 3 dollars 48 cents output. Flash drops to 3 cents input and about 28 cents output. Incredible value, yes. Automatic best model? Not even close.
Key facts
- DeepSeek V4 Preview introduces two MIT-licensed open-source models called Pro and Flash.
- The Pro model is described as having a 1 million token context window, 1.6 trillion total parameters, and 49 billion active parameters.
- The Flash model is described as having 284 billion total parameters and 13 billion active parameters.
- The source highlights a 1 million token context length as one of DeepSeek V4 Preview’s standout technical features.
- In the video’s comparison, Flash is described as sometimes performing better than Pro on certain prompts despite being the cheaper model.
Why it matters
A model can look dominant on paper through benchmarks, parameter counts, and context length, yet still underperform in practical coding and interface-generation tasks.
The Signal Brief
One sourced idea worth your attention, in your inbox. No noise.