DeepSeek V4 – almost on the frontier (simonwillison.net)

Article summary

DeepSeek has released its V4 series of AI models, including the DeepSeek-V4-Pro and DeepSeek-V4-Flash, which are 1 million token context Mixture of Experts models with a large number of parameters. The Pro model is the largest open weights model, larger than Kimi K2.6 and GLM-5.1. The models are available at a relatively low cost, with pricing starting at $0.14/million tokens input and $0.28/million tokens output for Flash. The models have been tested and compared to other models, with some users finding them to be competitive in terms of quality.

Main themes

• AI model releases
• Model pricing and cost
• Model performance and quality
• Trusted execution environments
• Private data and security
• Comparison to other models

What commenters say

• The DeepSeek V4 models are comparable in quality to other models like OpenAI's 5.4 and Opus 4.6, but at a lower cost.
• Some users find the models to be suitable for their needs, such as frontend development, despite not being state-of-the-art.
• The use of trusted execution environments, like Tinfoil, can provide a secure way to run models in the cloud while maintaining data privacy.
• Others are skeptical of the security of trusted execution environments and prefer to run models locally.
• The pricing of the models is seen as a major advantage, with some users calculating that they can get more tokens for their money compared to other models.
• There is disagreement about the value of the models, with some users finding them to be a good alternative to more expensive models and others seeing them as inferior.
• The test of generating a pelican image is seen as outdated and not a good measure of a model's capabilities.
• The models' performance is competitive with other models, but may not be the best choice for every use case.