Article · 2026-08-12 · 4 moments

DeepSeek V4 Pro 0813 (on OpenRouter)

DeepSeek V4 Pro 0813 (on OpenRouter) The latest DeepSeek Pro model is now available, via API only. I had to link to OpenRouter because DeepSeek don't have any obvious announcement page for their new model. I haven't been able to confirm if they plan to release the open weights, but given the weights are available for both April's deepseek-ai/DeepSeek-V4-Pro and July's deepseek-ai/DeepSeek-V4-Flash-0731 it seems likely. Interestingly I got very different looking pelicans for the three differen ✦ AI generated

Read original ↗

GUEST the author HOST Simon Willison

Prediction

DeepSeek will likely release the open weights for the new model.

The author cannot confirm DeepSeek's plans but considers an open-weights release likely because weights were released for the April DeepSeek-V4-Pro and July DeepSeek-V4-Flash-0731 models.

transcript

the author: I haven't been able to confirm if they plan to release the open weights, but given the weights are available for both April's deepseek-ai/DeepSeek-V4-Pro and July's deepseek-ai/DeepSeek-V4-Flash-0731 it seems likely.

Context◆

The latest DeepSeek Pro model is available only through an API, and OpenRouter is the practical link because DeepSeek has no obvious announcement page for it.

The author reports that the newest DeepSeek Pro release is API-only. They link to OpenRouter because DeepSeek does not appear to have published a clear announcement page.

transcript

the author: The latest DeepSeek Pro model is now available, via API only. I had to link to OpenRouter because DeepSeek don't have any obvious announcement page for their new model.

Example◆

The model produces substantially different-looking pelicans at low, medium, and high reasoning levels.

The author observes that the pelican images differ markedly depending on the selected reasoning level. They say this degree of variation is unusual compared with other models they have tested.

transcript

the author: Interestingly I got very different looking pelicans for the three different reasoning levels of low, medium, and high. I've not noticed this kind of difference from any other model:

Context

The model's benchmark results appear to have originated in the official DeepSeek WeChat group and were repeatedly reposted before appearing in a Hacker News table.

The author traces the benchmark information from the official DeepSeek WeChat group to a Reddit post, which moderators deleted as low-effort, and then to an ASCII-art table on Hacker News.

transcript

the author: In terms of benchmarks... as far as I can tell those were released to the Official DeepSeek WeChat Group, then copied and pasted into a post on Reddit which was deleted by the moderators for being "low-effort", then copied into this ASCII-art table on Hacker News.

Highlight slides

Related episodes

Article · Interconnects

Kimi K3: The open-weights escalation

→

Article · Import AI

Import AI 464: Fables writes GPU kernels; AI automation; and analog computation

→

Article · Latent Space

[AINews] Thinky's Inkling: 975B-A41B multimodal, new best American Apache 2.0 open model (with Inkling-Small, 276B-A12B)

→