Qwen3.8-27B-oQ3e-mtp

This model was quantized using oQ (oMLX v0.5.7) mixed-precision quantization.

Base model: Qwen/Qwen3.8-27B

Chat template: froggeric/Qwen-Fixed-Chat-Templates

Quantization details

• Model type: qwen3_5
• Bits: 3
• Group size: 64
• Format: MLX safetensors
• MTP: Preserved (mtp_num_hidden_layers: 1)
• Calibration: oQ3e (enhanced, imatrix-based; oqe_code_multilingual)

Environment

• Hardware: M5 MacBook Air 32GB
• Inference Framework: oMLX v0.5.7
• Max Concurrent Requests: 4
• Settings:
  • Thinking: Disabled
  • TurboQuant KV Cache: Enabled (4-bit)
  • Lightning 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	5195.3	63.23	197.1 tok/s	15.9 tok/s	13.246	87.0 tok/s	13.65 GB
pp4096/tg128	21265.9	68.07	192.6 tok/s	14.8 tok/s	29.956	141.0 tok/s	15.22 GB

Continuous Batching (pp1024 / tg128)

Batch	tg TPS	Speedup	pp TPS	pp TPS/req	TTFT(ms)	E2E(s)
1x	15.9 tok/s	1.00x	197.1 tok/s	197.1 tok/s	5195.3	13.246
2x	15.6 tok/s	0.98x	175.7 tok/s	87.8 tok/s	11654.5	28.022
4x	15.5 tok/s	0.97x	154.5 tok/s	38.6 tok/s	26158.3	59.608

Intelligence Benchmark

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

Benchmark	Accuracy	Correct	Total	Time(s)	Think
MMLU	70.0%	21	30	116.2	No
TRUTHFULQA	86.7%	26	30	39.1	No
GSM8K	96.7%	29	30	743.1	No
MATHQA	33.3%	10	30	219.6	No
HUMANEVAL	90.0%	27	30	748.3	No