DeepSeek Model Releases, Tracked

DeepSeek has officially open-sourced DeepSeek-V4-Flash-Vision-Exp, a 305-billion parameter native multimodal Mixture-of-Experts model published under the permissive MIT license on Hugging Face.

DeepSeek officially released V4-Pro-0813 on August 13, 2026 — the generally available version of its 1.6-trillion-parameter MoE model. It brings three-level thinking effort controls, a DSpark speculative decoding module, 87.9% on TerminalBench 2.1, and a peak/off-peak pricing structure that cuts API costs by 50% outside peak hours.

DeepSeek released V4-Flash-0731 on July 31, an updated version of their 284B MoE Flash model re-post-trained specifically for agentic and coding performance. TerminalBench 82.7 and NL2Repo 54.2 are the headline numbers. The Artificial Analysis Intelligence Index puts it 10 points above the previous Flash and competitively close to Pro-tier models. MIT-licensed weights and GGUF versions are already available.

DeepSeek V4 Pro (Expert Mode) one-shot a 696-line Three.js survival game that loaded sideways. Three follow-up fixes later — camera, food, terrain — the world worked, but you still can't switch to the axe or pick up food by hand. Full prompt, real fix log, and the unresolved bug inside.