Qwen 3.5 27B tops Gemma 4 on MI50

Qwen 3.5’s Multi-Token Prediction (MTP) architecture is a game-changer for throughput, though the "reasoning tax" remains a factor.

• –Qwen 3.5 27B achieves nearly double the output throughput of Gemma 4 31B, bolstered by a 5-token speculative decoding step with an 89.7% acceptance rate.
• –Gemma 4 31B shows substantially higher Time to First Token (43s vs 25s), suggesting less efficient prompt processing on the specific vllm-gfx906-mobydick stack.
• –The specialized "mobydick" fork is the key enabler for legacy gfx906 cards, using hybrid FP32/FP16 activations to bypass the lack of native BF16 support.
• –Despite the speed, Qwen’s integrated reasoning can lead to high token counts per request; for simple tasks, Gemma’s lower verbosity may still result in faster wall-clock completion.