API model IDs
deepseek-v4-flash / deepseek-v4-pro
Use deepseek-v4-flash as the default model ID in the DSFlashHub content stack; reserve deepseek-v4-pro for explicit escalation.
Model dossier
DeepSeek V4 Flash specs and Pro escalation notes
This page tracks source-backed model details only: IDs, context, output limit, reasoning modes, open weights, and architecture notes. The editorial priority is V4 Flash; Pro is documented as the escalation path.
Official details
V4 specification snapshot
DeepSeek API docs and the model card confirm the model IDs, context window, output limit, open-weight status, and parameter-scale positioning.
Context length
1M tokens
V4 Flash and V4 Pro both publish a 1M-token context window.
Max output
384K tokens
The model card lists a 384K maximum output length.
Reasoning modes
Non-think / Think / Think Max
Both Flash and Pro expose non-thinking and thinking modes; Think Max is reserved for the hardest reasoning cases.
Open weights
MIT licensed weights
The Hugging Face model card lists open weights and links to the V4 technical report.
Architecture
MoE + hybrid attention
V4 Pro is listed as 1.6T total / 49B active; V4 Flash is listed as 284B total / 13B active.
When to use Flash
V4 Flash is the lower-cost route for default chat, OpenClaw agent turns, retrieval augmentation, content processing, batch code explanation, and lower-risk tool calls.
When to use Pro
V4 Pro is the higher-reasoning route for difficult debugging, long-context repository review, incident analysis, and high-value reports. It should not replace every Flash request.