DeepSeek updates V4 Pro/Flash checkpoints, previews a vision variant and an agent harness
2026-08-13
DeepSeek updated its hosted checkpoints (deepseek-v4-flash to DeepSeek-V4-Flash-0731, deepseek-v4-pro to DeepSeek-V4-Pro-0813, both accessible under the same model names), released an experimental image-input variant (deepseek-v4-flash-vision-exp), and put a new 'DeepSeek Harness' agent harness into developer preview for agent-tool builders.
Significance 3: A real update to a tracked entity plus new agentic tooling, but no benchmark deltas are disclosed, so it can't be scored as a confirmed axis move.
Implications · machine-drafted, not owner judgment
DeepSeek shipping its own agent harness alongside checkpoint refreshes signals it's pushing beyond being just a cheap model host into full-stack agentic tooling, which matters for the bulk-cheap verdict's runner-up position — if the harness lowers the integration cost of using V4 for extraction/agentic work, that strengthens the case for a verdict flip. It's also incremental evidence for the open-weights-one-generation-behind thesis, though without disclosed scores it doesn't move the confidence number on its own.
- Independent benchmark numbers for DeepSeek-V4-Pro-0813 vs the prior checkpoint
- DeepSeek Harness reaching general availability or third-party adoption reports
Sources
- primary DeepSeek API docs — Your First API Call retrieved 2026-08-26
Models
Affected axes
Coding · Agentic / tool use · Cost