Models library

Qwen3 includes dense and mixture-of-experts models and supports reasoning and normal response modes. The Qwen3 family is publicly released under Apache 2.0.

Best launch choices: Qwen3-8B, Qwen3-30B-A3B and Qwen3-32B.

Best launch choice: DeepSeek-R1-Distill-Qwen-32B.

The 14B model is better for affordable single-GPU jobs. The 32B version is better positioned as the standard reasoning product, while the 70B version should be a premium RTX PRO 6000 deployment.

Qwen3-Coder-30B-A3B is the practical single-GPU version, whereas the 480B model is intended for much larger deployments. Qwen3-Coder-Next is an 80B-parameter mixture-of-experts model activating approximately 3B parameters during inference, but its total model memory still makes the RTX PRO 6000 the safer deployment choice.

Best launch choice: Qwen3-Coder-30B-A3B-Instruct FP8.

Qwen3-VL supports text, images and video, with dense and mixture-of-experts variants. The model family supports long multimodal contexts, making it useful for document collections and video analysis.

Best launch choices: Qwen3-VL-8B FP8 and Qwen3-VL-30B-A3B.

Best hardware: RTX 5090. Image-generation workloads usually do not need 96 GB. The RTX 5090 gives customers high performance without charging them for unused RTX PRO 6000 memory. Use the RTX PRO 6000 for large batches, multiple simultaneous workflows or unusually large pipelines. Offer these through ComfyUI, including persistent storage for customer checkpoints, LoRAs, ControlNets and workflows.

Best launch choice: Wan2.2 TI2V-5B on RTX 5090.

Use the RTX PRO 6000 for longer videos, higher resolutions, heavier ComfyUI graphs, larger batches or concurrent users. The Intel B70 may eventually be viable, but video-generation pipelines commonly depend on CUDA-specific components, so it should not be your default B70 product until fully validated.

Best hardware: Intel Arc Pro B70. Whisper does not require 96 GB of VRAM. A B70 is a better economic match, provided your selected inference runtime is stable on Intel. The RTX 5090 is the safer fallback for maximum software compatibility. For shared transcription, sell by audio minute rather than reserving an entire GPU.