[throwa356262]

Deepseek v4 via deepseek themselves is significantly cheaper.

Because (1) Huawei collab and (2) vLLM etc dont implement half of the inference optimisations deepseek proposed in their paper.