DeepSeek V4: Chinese AI Challenges the Market with Its Pricing

The Chinese artificial intelligence laboratory DeepSeek has recently unveiled its new models from the V4 series, marking a significant advancement in the field of large-scale language models. The two models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, are previews that stand out due to their size and attractive cost.

The latest model from the lab, DeepSeek V3.2, along with its special version, was released last December. Today, DeepSeek launches the first models of the highly anticipated V4 series. These models, DeepSeek-V4-Pro and DeepSeek-V4-Flash, are designed as Mixtures of Experts with an impressive context of 1 million tokens. The Pro model, with its 1.6 trillion parameters, of which 49 billion are active, positions itself as the largest open-weight model currently available, surpassing competitors such as Kimi K2.6 (1.1T) and GLM-5.1 (754B), and more than twice the size of DeepSeek V3.2 (685B). The Flash model, on the other hand, has 284 billion parameters, with 13 billion active. These models use the standard MIT license.

In terms of storage, the Pro model requires 865 GB on the Hugging Face platform, while the Flash model requires 160 GB. The author hopes that a slightly quantized Flash model could run on a MacBook Pro M5 with 128 GB. It is anticipated that the Pro model could operate by streaming only the necessary active experts from the disk.

To test these models, the author used OpenRouter, installing the llm-openrouter module. Here is an example command used to generate an SVG of a pelican riding a bicycle:

llm install llm-openrouter
llm openrouter refresh
llm -m openrouter/deepseek/deepseek-v4-pro 'Generate an SVG of a pelican riding a bicycle'

The results for DeepSeek-V4-Flash and DeepSeek-V4-Pro were compared with those obtained previously with DeepSeek V3.2 in December, V3.1 in August, and V3-0324 in March 2025. While the generated pelicans were of good quality, the most remarkable aspect of these new models is their cost.

DeepSeek offers extremely competitive pricing: $0.14/million tokens for input and $0.28/million tokens for output for the Flash model, and $1.74/million for input and $3.48/million for output for the Pro model. These prices challenge those of leading models from Gemini, OpenAI, and Anthropic. For example, DeepSeek-V4-Flash is the cheapest among small models, even beating OpenAI's GPT-5.4 Nano, while DeepSeek-V4-Pro is the cheapest among large leading models.

A note from DeepSeek's document explains why they can offer these models at such a low price: they focused on efficiency with this version, particularly for longer context prompts. In the scenario of a 1 million token context, even DeepSeek-V4-Pro, which has a larger number of active parameters, only reaches 27% of the FLOPs at a single token (measured in equivalent FP8 FLOPs) and 10% of the KV cache size compared to DeepSeek-V3.2. Additionally, DeepSeek-V4-Flash, with its smaller number of active parameters, pushes efficiency even further: within a 1 million token context, it only reaches 10% of the FLOPs at a single token and 7% of the KV cache size compared to DeepSeek-V3.2.

DeepSeek's self-reported benchmarks in their document show that their Pro model is competitive with these other leading models, although with this note: through the expansion of reasoning tokens, DeepSeek-V4-Pro-Max demonstrates superior performance compared to GPT-5.2 and Gemini-3.0-Pro on standard reasoning benchmarks. Nevertheless, its performance is slightly lower than that of GPT-5.4 and Gemini-3.1-Pro, suggesting a development trajectory that lags behind by about 3 to 6 months compared to leading models.

Attention now turns to future quantized versions from the Unsloth team, which could further enhance the accessibility and efficiency of these models on personal machines. The author is keeping an eye on huggingface.co/unsloth/models, expecting that the Unsloth team will soon have a series of quantized versions. It will be very interesting to see how this Flash model performs on their own machine.