Summary & Key Takeaways ​

• DeepSeek has released its new V4 Pro 0813 large language model.
• The model is currently available for use exclusively through API access.
• Official announcements and benchmark data for the release are notably scarce.
• Simon Willison observed unique differences in image generation based on reasoning levels.
• There's speculation about the eventual open-sourcing of the model weights.

Our Commentary ​

DeepSeek's new model is out, but the release process sounds like a wild goose chase. Benchmarks on WeChat, then Reddit, then Hacker News? That's... something. I appreciate Simon Willison cutting through the noise to give us the lowdown. The pelican observation is genuinely fascinating; it makes me wonder what other subtle differences exist across models.