Overview
DeepSeek-V4-Flash-Base is the base checkpoint of the DeepSeek-V4-Flash mixture-of-experts language model. The DeepSeek-V4 preview series was announced on April 24, 2026. DeepSeek lists this checkpoint separately from the post-trained DeepSeek-V4-Flash model. (api-docs.deepseek.com)
Capabilities and uses
- Architecture: 284 billion total parameters, with 13 billion activated; mixed FP8 precision.
- Context: Up to 1 million tokens.
- Capabilities: DeepSeek reports base-model evaluations covering knowledge, language and reasoning, code, mathematics, and long-context tasks.
- Uses: Research, adaptation or fine-tuning, and local deployment where suitable infrastructure is available. (huggingface.co)
Limitations and availability
This is a base checkpoint, not the instruction-tuned Flash model; the latter's chat modes and API features should not be assumed to apply to it. The weights require substantial deployment resources: the official repository lists approximately 295 GB of files. The checkpoint is distributed under the MIT License. (huggingface.co)
DeepSeek's current pricing pages do not list a separate API offering for this base checkpoint. The older deepseek-v4-flash API name now routes to a newer post-trained model, not this base checkpoint. (api-docs.deepseek.com)
Pricing
Unverified for this exact base checkpoint: No separate official CNY or USD input-token or output-token API price was found in the domestic and international pricing pages checked on November 16, 2026. Prices for other Flash versions are not prices for this checkpoint. Self-hosting can incur infrastructure costs.