DeepSeek-V4-Flash-oQ2.5e-mtp

This model was quantized using oQ (oMLX v0.5.0.dev1) mixed-precision quantization.

Quantization details

• Model type: deepseek_v4
• Bits: 2
• Group size: 64
• Format: MLX safetensors

Accuracy Benchmark

Runs used the oMLX standard accuracy benchmark harness with greedy decoding (temperature=0, top_p=1), thinking disabled, and batch size 32.

Type	Model	Size	Avg	MMLU 1000	Winogrande 1267	HumanEval 164	MBPP 300	MMLU-Pro 1000	KMMLU 500
Original precision	DeepSeek V4 Flash (MLX engine)	148.7 GiB	78.09%	821/1000 82.10%	926/1267 73.09%	148/164 90.24%	249/300 83.00%	677/1000 67.70%	362/500 72.40%
MLX oQe	DeepSeek-V4-Flash-oQ2.5e	103.1 GiB	74.56%	784/1000 78.40%	867/1267 68.43%	146/164 89.02%	234/300 78.00%	665/1000 66.50%	335/500 67.00%
GGUF	Q4K	153.3 GiB	78.64%	820/1000 82.00%	946/1267 74.66%	147/164 89.63%	250/300 83.33%	682/1000 68.20%	370/500 74.00%
GGUF	Q4K+IQ2XXS	90.9 GiB	69.85%	596/1000 59.60%	1024/1267 80.82%	146/164 89.02%	221/300 73.67%	624/1000 62.40%	268/500 53.60%