KAT-Coder-V2.5-Dev-oQ2e-mtp

This model was quantized using oQ (oMLX v0.5.4) mixed-precision quantization.

• Base model: Kwaipilot/KAT-Coder-V2.5-Dev
• MTP head: Qwen/Qwen3.6-35B-A3B (original MTP head grafted)
• Chat template: Qwen Fixed Chat Template (v21) — fixes KV cache optimization, thinking mode control, tool calling, and agent loop issues. Original backed up as chat_template.jinja.bak.

Quantization details

• Model type: qwen3_5_moe
• Bits: 2
• Group size: 64
• Format: MLX safetensors
• MTP: Grafted from Qwen3.6-35B-A3B (bf16 original)
• Calibration: oQ2e (enhanced, imatrix-based)

Environment

• Hardware: M5 MacBook Air 32GB
• Inference Framework: oMLX v0.5.4
• Max Concurrent Requests: 4
• Settings:
  • Thinking: Disabled
  • 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	1119.8	17.25	914.4 tok/s	58.4 tok/s	3.330	346.0 tok/s	14.20 GB
pp4096/tg128	4040.2	21.25	1013.8 tok/s	47.4 tok/s	6.754	625.4 tok/s	14.94 GB

Continuous Batching (pp1024 / tg128)

Batch	tg TPS	Speedup	pp TPS	pp TPS/req	TTFT(ms)	E2E(s)
1x	58.4 tok/s	1.00x	914.4 tok/s	914.4 tok/s	1119.8	3.330
2x	79.3 tok/s	1.36x	825.0 tok/s	412.5 tok/s	2482.5	5.711
4x	112.2 tok/s	1.92x	827.0 tok/s	206.8 tok/s	4815.5	9.517

Intelligence Benchmark

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

Benchmark	Accuracy	Correct	Total	Time(s)	Think
MMLU	73.3%	22	30	36.3	No
TRUTHFULQA	86.7%	26	30	15.5	No
GSM8K	96.7%	29	30	76.7	No
MATHQA	13.3%	4	30	76.2	No
HUMANEVAL	90.0%	27	30	118.5	No