Direct answer
There is no single provider-independent context limit to assume. For DeepSeek V3.2 on Amazon Bedrock, AWS documents a 164K-token context window and a maximum output of 8K tokens. The Bedrock model ID is deepseek.v3.2. These figures apply to AWS’s hosted deployment—not automatically to DeepSeek’s own API, self-hosted weights, or another provider. The AWS model card gives a launch date of December 1, 2025, and currently lists the model as active.
What about DeepSeek’s own API?
DeepSeek’s November 30, 2025 release announcement said V3.2 was available on its API and used the same usage pattern as V3.2-Exp, but it did not specify a numeric context limit or output cap for standard V3.2. The current DeepSeek API quick-start lists deepseek-flash and deepseek-v4-pro; it does not list V3.2. So the AWS figure is a concrete production-API specification, but it does not establish what limit applied to DeepSeek’s own API historically or whether that exact model is currently offered there.
DeepSeek’s API model-list schema defines context_window as the total capacity counting input and output tokens, with max_output_tokens as a separate maximum-output field. Those are schema definitions, not V3.2-specific values; don’t apply them to AWS without checking AWS’s own request rules.
Check whether a request fits
For the Bedrock deployment, use the documented 164K context-window figure and keep the requested response at or below the 8K output cap. Estimate the tokens in the prompt and conversation history, then account for the intended response under the provider’s request rules. AWS’s model card lists the context window and output cap separately; confirm the endpoint’s token-counting behavior rather than assuming the full 164K is available for input alone.
For a different host, check that provider’s exact model ID, context limit, and output cap. Don’t transfer the Bedrock limit to another API or deployment.
Sparse attention is not a context limit
DeepSeek describes DeepSeek Sparse Attention (DSA) as improving long-context efficiency. That describes an efficiency goal; it does not, by itself, establish a supported token maximum.
Documentation checked November 10, 2026.
Sources
DeepSeek V3.2 — Amazon Bedrock model card
Your First API Call — DeepSeek API Docs