inferencerlabs/MiMo-V2.5-Pro-MLX-4.3bit-INF

See MiMo-V2.5-Pro in action - demonstration videos

Tested with an M3 Ultra 512 GiB and M4 Max 128 GiB RAM using Inferencer app v1.11.2 distributed compute

• Distributed inference: ~13 tokens/s @ 1000 tokens ~450 GiB / ~67 GiB (debug build)

Q4.3-INF uses the data-agnostic INF method tuned to yield maximum general accuracy within a distributed 640 GiB memory budget

This quantization level represents the minimum at which the model maintained acceptable response coherence, as more aggressive quantization resulted in significant degradation.

Additionally, the perplexity of this quantization has not been directly compared against the base model due to resource and time constraints. For general guidance, the evaluations below reference a similarly sized model (Kimi K2.6). However, please note that Kimi K2.6 uses quantization-aware training (QAT), so these results are not directly comparable and should be treated as context only.

Quantization (bpw)	Perplexity	Token Accuracy	Missed Divergence
Q3.5	1.1328125	94.92%	42.71%
Q3.5-INF	1.078125	96.67%	22.04%
Q3.6	1.1484375	94.72%	48.72%
Q4.2-INF	1.0546875	99.02%	13.73%
Base	Untested	100%	0.000%

• Perplexity: Measures the confidence for predicting base tokens (lower is better)
• Token Accuracy: The percentage of correctly generated base tokens
• Missed Divergence: Measures severity of misses; how much the token was missed by

Quantized with a modified version of MLX

For more details see our demonstration videos or visit MiMo-V2.5-Pro.

Disclaimer

We are not the creator, originator, or owner of any model listed. Each model is created and provided by third parties. Models may not always be accurate or contextually appropriate. You are responsible for verifying the information before making important decisions. We are not liable for any damages, losses, or issues arising from its use, including data loss or inaccuracies in AI-generated content.