DeepSeek · Model release
DeepSeek V4-Flash-0731: why the post-training update matters
The official V4-Flash build keeps the architecture but refreshes post-training for agents, coding, and tool use at unusually low API rates.
What changed
- The 0731 build moved the API into public beta and published updated weights.
- The architecture stayed broadly the same; the claimed gains come from post-training and serving changes.
- The release supports a one-million-token context and a low-cost hosted route.
Best fit
- Cost-sensitive agents
- Open-weight experiments
- High-volume coding
What to be careful about
- Many benchmark numbers are vendor-reported and use a DeepSeek harness.
- Open weights can still require datacentre-class memory.
Dataset snapshot: DeepSeek V4-Flash is also represented in the checked comparison dataset. Prices and benchmark figures carry their own field dates.