Rapid-MLX on the Same Qwen3.6 35B A3B Model: Fast First Token, Slower Sustained Run

Rapid-MLX was cloned into /Users/jason/Developer/projects/tool-rapid-mlx and installed from source in an editable virtualenv. LM Studio was unloaded with lms unload --all before the run. Rapid-MLX served ~/.lmstudio/models/mlx-community/Qwen3.6-35B-A3B-4bit on localhost:8000 with --served-model-name qwen3.6-35b-a3b, --max-concurrent-requests 1, --max-num-seqs 1, --no-thinking, and --no-mllm. The --no-mllm flag was needed because the LM Studio model directory includes processor files and the text-only Rapid-MLX install does not include mlx-vlm.

Rapid-MLX produced 6144 completion tokens in 87.21s at 70.45 tok/s with a 0.18s time to first token. That first token was excellent, but sustained throughput trailed the prior LM Studio result at 88.55 tok/s and the prior oMLX result at 87.19 tok/s. In the shared-prefix benchmark, Request A TTFT was 3.21s and Request B TTFT was 3.23s, so Request B was effectively flat rather than faster. Rapid-MLX did not report cached token counts through the API in this run, and the server logs showed cache misses for both long prefix requests.