KAT-Coder-V2.5-Dev-oQ2e

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

• Base model: Kwaipilot/KAT-Coder-V2.5-Dev
• 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
• Calibration: oQ2e (enhanced, imatrix-based)

Environment

• Hardware: M5 MacBook Air 32GB
• Inference Framework: oMLX v0.5.4
• Max Concurrent Requests: 4
• Settings:
  • Thinking: Disabled
  • TurboQuant KV Cache: Enabled (4-bit)

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	1103.1	20.35	928.3 tok/s	49.5 tok/s	3.700	311.4 tok/s	12.62 GB
pp4096/tg128	3794.9	21.39	1079.3 tok/s	47.1 tok/s	6.530	646.9 tok/s	13.34 GB

Continuous Batching (pp1024 / tg128)

Batch	tg TPS	Speedup	pp TPS	pp TPS/req	TTFT(ms)	E2E(s)
1x	49.5 tok/s	1.00x	928.3 tok/s	928.3 tok/s	1103.1	3.700
2x	67.5 tok/s	1.36x	836.8 tok/s	418.4 tok/s	2447.4	6.238
4x	98.7 tok/s	1.99x	825.1 tok/s	206.3 tok/s	4822.0	10.150

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	39.5	No
TRUTHFULQA	86.7%	26	30	16.3	No
GSM8K	93.3%	28	30	99.4	No
MATHQA	20.0%	6	30	76.1	No
HUMANEVAL	86.7%	26	30	109.7	No