deepseek-v4.1-flash in the catalog and in GET /v1/models, sent as model; its current version is deepseek-v4.1-flash-1. Its page on the site, with its rate history: DeepSeek: DeepSeek V4.1 Flash. DeepSeek V4.1 Flash is a sparse mixture-of-experts model from DeepSeek with 552B total parameters (8B active per input token), Engram conditional memory, and a 1M-token context window. It is suited for coding, agent workflows, and high-volume tool use, with a controllable reasoning effort.Reasoning is supported and enabled by default.
Fig. 1, sheet 06Model and task guides
What it takes and returns
Field
Value
Type and task
language, task chat
License and usage conditions
apache
Field
Type and task
Value
language, task chat
Field
License and usage conditions
Value
apache
Limits
Limit
Value
context_tokens
1,048,576
max_output_tokens
1,048,576
Limit
context_tokens
Value
1,048,576
Limit
max_output_tokens
Value
1,048,576
Pools
Pool
Lanes
Precision
Certified pool
Shared
fp4
Pool
Certified pool
Lanes
Shared
Precision
fp4
Rates
On demand, in US dollars per million tokens
Pool
Input
Cached input
Output
Certified pool
$0.098
Billed as input
$0.70
Pool
Certified pool
Input
$0.098
Cached input
Billed as input
Output
$0.70
Request parameters
In the Certified pool it takes frequency_penalty, logit_bias, presence_penalty, stop, temperature, tools and top_p. A request carrying any other parameter is answered 400 unsupported_parameter, its param naming it, and nothing is charged (Request parameters).
Always taken as well: model, messages, stream, stream_options, max_tokens, max_completion_tokens, user, metadata, safety_identifier, store and prompt_cache_key and the hips object.