Qwen3.5-9B-oQ4e-mtp

This model was quantized using oQ (oMLX v0.5.1) mixed-precision quantization.

Quantization details

• Model type: qwen3_5
• Bits: 4
• Group size: 64
• Format: MLX safetensors

Environment

• Hardware: M5 MacBook Air 32GB
• Inference Framework: oMLX v0.5.1
• Settings:
  • Thinking: Disabled
  • Chat template parameter: enable_thinking=false (forced)
  • TurboQuant KV Cache: Disabled (if enable: reduces speed)
  • Native MTP: Enabled (key speed improvement)

Performance Benchmarks

Note: Results are for reference only and may vary depending on hardware, software configuration, and workload.

Single Request Results

Test	TTFT(ms)	TPOT(ms)	pp TPS	tg TPS	E2E(s)	Throughput	Peak Mem
pp1024/tg128	1591.5	23.16	643.4 tok/s	43.5 tok/s	4.556	252.8 tok/s	5.96 GB
pp4096/tg128	5941.8	23.44	689.4 tok/s	43.0 tok/s	8.946	472.2 tok/s	6.58 GB

Continuous Batching (pp1024 / tg128)

Batch	tg TPS	Speedup	pp TPS	pp TPS/req	TTFT(ms)	E2E(s)
1x	43.5 tok/s	1.00x	643.4 tok/s	643.4 tok/s	1591.5	4.556
2x	50.7 tok/s	1.17x	641.2 tok/s	320.6 tok/s	3193.8	8.248
4x	81.7 tok/s	1.88x	638.1 tok/s	159.5 tok/s	6263.5	12.683

Intelligence Benchmark

Note: Each benchmark round tests only 30 questions. Results are for reference only.

Benchmark	Accuracy	Correct	Total	Time(s)	Think
MMLU	76.7%	23	30	34.0	No
TRUTHFULQA	86.7%	26	30	12.4	No
GSM8K	90.0%	27	30	262.0	No
MATHQA	26.7%	8	30	20.5	No
HUMANEVAL	86.7%	26	30	144.9	No

Summary

Benchmark	Accuracy
MMLU	76.7%
TRUTHFULQA	86.7%
GSM8K	90.0%
MATHQA	26.7%
HUMANEVAL	86.7%
Average	73.3%