Benchmark

Model Alias	Actual Model	Tokens/s	Notes
qwen3-coder:30b	-	31.84	-
qwen3.5:9b	qwen3.5:9b-q4_K_M	15	Tool-calling accuracy has been improved compared to Qwen3.
qwen3:35b-a3b (262k)	qwen3.5:35b-a3b-q4_K_M	25	Manually adjust the context length from 8K to 262K; tool calling capability is stable; VRAM usage remains consistently between 24.4G and 30G.