inferencerlabs/Kimi-K2.6-MLX-3.5bit-INF

See Kimi-K2.6 MLX in action - demonstration video

Tested on an M3 Ultra 512 GiB RAM using Inferencer app v1.11

• Text inference: ~21.04 tokens/s @ 1000 tokens ~423.61 GiB (debug build)
• 2x Batched inference: ~30.1 tokens/s
• Vision inference: ~18.28 tokens/s ~426.53 GiB

Q3.5-INF uses the data-agnostic INF method tuned to yield maximum general accuracy within a 512 GiB memory budget

Due to system memory and time constraints, the base generation was inferenced directly from storage and limited to the first 512 tokens for our coding tests. To address this, generation was configured to output complex functionality from the start rather than incremental scaffolding. While this setup differs from typical usage, it stresses early-token accuracy and the observed trends appear consistent with those seen in larger-scale evaluations of other models. These figures may be updated in future with extended evaluations.

Quantization (bpw)	Perplexity	Token Accuracy	Missed Divergence	Size
Q3.5	1.1328125	94.92%	42.71%	450.19 GB
Q3.5-INF	1.078125	96.67%	22.04%	455.68 GB
Q3.6	1.1484375	94.72%	48.72%	470.99 GB
Base	Untested	100%	0.000%	658.59 GB

• 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 video or visit Kimi-K2.6.

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.