Choose the exact Qwen generation and variant
Two distinct starting points are qwen/qwen3.8-27b-uncensored and huihui-ai/Qwen2.5-32B-Instruct-abliterated. They differ in generation, size and modification; the labels are not interchangeable. Use their individual model pages for the exact context, endpoint and current customer rates.
A DeepSeek-R1-Distill-Qwen model has Qwen in its name but belongs to the DeepSeek comparison in these guides. Do not assume that a distilled reasoning variant and a Qwen instruct variant use the same prompting conventions or produce the same answer style.
Build a small evaluation before choosing
For a multilingual assistant, use prompts in the actual languages your customers use and ask a reviewer to check fluency, meaning and instruction following. For coding assistance, compare a small function, a failing test explanation and a change to an existing snippet. These are suggested evaluations, not claimed benchmark results for the listed variants.
Keep a fixed output budget and inspect both correctness and response length. A verbose answer can consume more output tokens even when the listed price is lower. Use the live table to estimate your workload with separate input and output rates.
Use one integration and swap only the model ID
The Qwen entries below currently advertise chat completions. Keep https://kearnix.com/v1 and your server-side Kearnix key unchanged when comparing variants; replace only KEARNIX_MODEL with the full ID. The Python and Node.js guides use that environment variable so a comparison does not require rewriting the client.
The catalog currently does not confirm tool use, image input or JSON schema support for these entries. Verify a required feature on the exact variant rather than inheriting claims from another Qwen release. Availability and prices below update from the public catalog.