Skip to content
4ALL API Docs
Main Site News Console

deepseek-ai/DeepSeek-

· Simon Willison Translated
技巧LLM

deepseek-ai/DeepSeek-V4-Flash-0731

The latest version of the DeepSeek V4 series, featuring "significantly enhanced agent capabilities." It has 304 billion parameters and takes up 167GB on Hugging Face—but it looks like its performance far exceeds its size.

Artificial Analysis ranks it ahead of MiniMax M3—a 428-billion-parameter model. With an input price of $0.14 per million tokens and an output price of $0.27 per million tokens, this means it is currently perhaps the most cost-effective intelligent model on the market. On the Intelligence Index vs. Cost per Intelligence Index Task chart, its performance looks very impressive:

A scatter plot from Artificial Analysis with axes 'Artificial Analysis Intelligence Index' (20 to 65) and 'Cost per Task (USD, Log Scale)' ($0.02 to $3). There is a green 'Most attractive quadrant' box in the top left, and a dashed 'Pareto line'. DeepSeek V4 Flash 0731 (max) is highlighted in dark blue, positioned at approximately $0.028 and an intelligence score of 50, sitting alone on the far left edge of the green quadrant where the Pareto line curves sharply upward. Models with similar or lower intelligence levels, such as MiniMax-M3, Kimi K3 (low), GLM-5.1, and Kimi K2.6, cost ten times more; while stronger models (Grok 4.5, Gemini 3.6 Flash, GLM-5.2, Kimi K3, Claude Opus 5, Claude Fable 5, GPT-5.6 Sol) are all located further to the right, with costs per task ranging from $0.4 to $3.

Using the default reasoning level via OpenRouter, I got a disappointing pelican:

A flat vector illustration: a white pelican with a long neck and a large orange pouch, hovering above a jumbled blue and orange bicycle, set against a dark gray road with dashed white lane markings. The bicycle is drawn incorrectly: the wheels are just orange arcs without rims or spokes, the frame tubes are disconnected and floating, and the handlebars are not attached to anything. The background is light blue, with a yellow sun and white clouds in the top left, and gray speed lines on the left suggesting motion.

But when I turned the reasoning level up to high, I got a much better result:

llm -m openrouter/deepseek/deepseek-v4-flash-0731 -t pelican -o reasoning_effort high

A flat vector illustration: a white pelican riding a bicycle to the right, against a pink background with a lighter pink circle behind it. The pelican is holding the handlebars with its wings, with one orange foot on the pedal, and a small blue fish peeking out from the corner of its large orange pouch. The bicycle frame is red, blue, and orange, with dark tires, and gray speed lines trailing behind to show motion.