DeepSeek V4

DeepSeek V4 was released suddenly without warning on Hugging Face, with an official announcement on X following shortly after.

The model is open-source, with weights available on Hugging Face for anyone to download and run locally.

The author does not have a machine capable of running the V4 Pro model locally.

DeepSeek V4 Flash has 158 billion parameters and is a good candidate for hobbyist/enthusiast local setups.

DeepSeek V4 Pro is a 1.66 trillion parameter mixture-of-experts model with 49 billion active parameters and a 1 million token context length.

Benchmarks compare V4 to older Opus 4.6, GPT 5.4, and Gemini 3.1 Pro High, showing comparable or superior performance in some tests.

V4 models are significantly more efficient than V3.2 in computational cost and memory usage.

The model has a maximum output length of 384K tokens.

API pricing is $1.74 per million input tokens and $348 per million output tokens for V4 Pro, while GPT-5.5 costs $5 per million input and $30 per million output.

In the browser OS test, V4 Pro generated a 2700-line webOS with a functional right-click menu, correct clock, and Windows-style start menu.

The GTA clone included buildings, trees, grid streets, and an NPC car but lacked a wanted system or pickups, with the car spawning inside a building.

Initial calculator, notepad, terminal, and file explorer apps did not load; the issues were resolved by fixing the code in Open Code.

The terminal could fully control the GUI environment, including moving windows and changing backgrounds via commands.

The subway scene featured arched ceilings, vending machines, and detailed tiles, but movement was inverted and stairs lacked collision.

The flight combat simulator had an impressive plane selection screen with hover effects, but no enemies were visible during gameplay.

The drum kit simulation (via chat.deepseek.com) produced good model details, stands, and an autoplay feature; the hip-hop beat was notably hard.

The 3D printer simulation with thinking mode showed a coreXY printer with correct filament path and stepped bed movement; without thinking, it was basic pancake stacking.

The C++ skateboarding game had a boardwalk aesthetic with buildings and many humanoids, but glitches like falling through the map undermined its potential.

The front-end AI analytics dashboard was clean and aesthetically pleasing but had a non-functional search bar.

Overall, the V4 Pro preview model produced satisfactory but not mind-blowing results, feeling like a preview release.

Total testing cost was $3.68, and the model's API pricing is significantly cheaper than competitors like GPT-5.5 and Opus.

DeepSeek V4 is the largest ever open-weight model release.

The Flash version's size makes it very exciting for local AI enthusiasts.