Qwen3.8-27B-oQ2e-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: 2
• Group size: 64
• Format: MLX safetensors
• MTP: Preserved (mtp_num_hidden_layers: 1)
• Calibration: oQ2e (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	6074.9	67.08	168.6 tok/s	15.0 tok/s	14.615	78.8 tok/s	11.71 GB
pp4096/tg128	44281.6	96.54	92.5 tok/s	10.4 tok/s	56.590	74.6 tok/s	13.21 GB

Continuous Batching (pp1024 / tg128)

Batch	tg TPS	Speedup	pp TPS	pp TPS/req	TTFT(ms)	E2E(s)
1x	15.0 tok/s	1.00x	168.6 tok/s	168.6 tok/s	6074.9	14.615
2x	11.2 tok/s	0.75x	142.4 tok/s	71.2 tok/s	14382.5	37.157
4x	13.1 tok/s	0.87x	129.4 tok/s	32.4 tok/s	31210.7	70.791

Intelligence Benchmark

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

Benchmark	Accuracy	Correct	Total	Time(s)	Think
MMLU	56.7%	17	30	189.5	No
TRUTHFULQA	63.3%	19	30	77.2	No
GSM8K	80.0%	24	30	649.4	No
MATHQA	23.3%	7	30	396.5	No
HUMANEVAL	83.3%	25	30	1468.4	No