NInfer-2080ti-22g: Engine for Qwen on a Single Modified RTX 2080 Ti Now Runs on Windows
Version 0.3.0 of NInfer has been released — an engine for running Qwen3.6/8-27B on a single RTX 2080 Ti with 22 GB. Windows is now supported, size — 0.28 GB. We break down what it is and who it's for.
NInfer-2080ti-22g is an engine for running Qwen3.6/8-27B models (only these) on a single modified RTX 2080 Ti graphics card with 22 GB of video memory. Version 0.3.0 was released on October 30, 2026. The main change is native Windows support. Previously, Linux, WSL2, or Docker were required. The distribution size is 0.28 GB. The price is not disclosed.
The engine uses only video memory, without offloading to system RAM. Therefore, models like 35B-A3B are not supported. Fine-tunes and abliterations (models with altered behavior) are supported, but they must be converted via the .ninfer checkpoint converter. Working with the tool looks like this: you install NInfer, convert the required Qwen3.6/8-27B checkpoint to the .ninfer format, then run inference. Everything is done locally, without the cloud. Example: you have an RTX 2080 Ti with 22 GB (modified), and you want to run Qwen3.6-8B. You convert the checkpoint, launch the engine — it loads the weights into video memory and generates responses. System RAM is not used, so the amount of video memory is the only limitation.
How does it differ from alternatives? The closest analogues are llama.cpp, ExLlama, vLLM. They support a broader range of models and architectures, but often require either more video memory or are able to offload layers to RAM. NInfer, on the other hand, is optimized strictly for Qwen3.6/8-27B and only for video memory. This provides simplicity and predictability, but narrows the scope of application. In addition, NInfer requires a modified card with 22 GB — a standard RTX 2080 Ti has 11 GB. The modification is performed in hardware, and its cost and availability are not disclosed here.
Who it's not for: owners of standard RTX 2080 Ti (11 GB) — there won't be enough memory. Those who want to run other models (Llama, Mistral, etc.) — the engine does not support them. Those who have no way to modify the graphics card. It is also not suitable if you need offloading to RAM for large models — there is none here. And those who are not ready to convert checkpoints manually.
Bottom line: NInfer-2080ti-22g v0.3.0 is a highly specialized tool for enthusiasts with a modified RTX 2080 Ti who want to run Qwen3.6/8-27B locally on Windows. It does not try to be universal, but solves a specific task. If you have the right hardware and the right model — it's worth trying. In all other cases, look at more flexible alternatives.