Separate Flash revisions from R1 distilled models
The catalog includes deepseek/deepseek-v4.1-flash-uncensored and dated Flash variants ending in -0910-uncensored and -1017-uncensored. Treat each full identifier as a separate deployment choice. A date suffix is not a compatibility guarantee, and the listed context windows can differ.
The huihui-ai/DeepSeek-R1-Distill-Qwen-32B-abliterated and huihui-ai/DeepSeek-R1-Distill-Llama-70B-abliterated entries are different distilled variants. Their names identify different base families and sizes; do not treat them as aliases for Flash or assume identical context and prompting behavior.
Evaluate explanations, correctness and cost together
For a reasoning workload, choose questions with answers you can independently check. Compare final correctness, adherence to your response format and output tokens. A longer explanation is not evidence of a more accurate answer. For coding, include a small deterministic test case rather than judging only how convincing the explanation sounds.
If you switch from a dated Flash revision to another variant, repeat a representative evaluation and check the context limit first. The table below lists current customer prices rather than a fixed cheapest-DeepSeek claim. Input and output rates must both be included when comparing long responses.
Integrate through the documented endpoint
The listed variants currently expose /v1/chat/completions through Kearnix. Set KEARNIX_MODEL to the full identifier and use the Python or Node.js setup below. Keep initial requests short and inspect usage before sending long conversation histories.
The current catalog does not establish support for agent tool calls, reasoning-control parameters or structured output. Do not pass parameters from a different DeepSeek product without checking this exact model. Kearnix Cloud is the uncensored service label; it does not change the model's developer attribution.