Bijan Bowen tests the released DeepSeek V4 Flash after benchmark gains that place it closer to much larger coding models. The mixture-of-experts model has 284 billion total parameters but activates about 13 billion at a time, making quantized local deployment plausible on high-memory enthusiast hardware while its hosted API remains inexpensive.
The model produced a functional browser-style operating system with working windows, files and applications, then generated several interactive 3D games and scenes. It handled unfamiliar C++ tasks, a printable motor enclosure and a multi-era city simulation better than expected for a Flash-class model, often producing polished interfaces and surprisingly complete interactions.
The strongest recurring weakness was spatial consistency. Some 3D objects were oriented incorrectly, vehicles and characters became distorted, and game logic occasionally failed or required repair. A few tests were also hard to interpret cleanly because earlier generated work may have influenced later attempts.
Bijan Bowen concludes that the combination of capability, low API cost and potential local deployment makes DeepSeek V4 Flash unusually competitive for its size. The results also increase expectations for the larger DeepSeek V4 Pro, provided its gains carry over without the same spatial and reliability problems.
Watch the original on YouTube