MLX implementation NIAH (@Prince_Canuma, Apple Silicon)

success
0.14
1/5
Overview Experiments 96 Forks 3 Resources 36 Benchmarks 2 Broadcasts 3 Related
Consensus Metrics
niah_2_5bit_score 6 (n=1, σ=0)
niah_2_5bit_total 6 (n=1, σ=0)
niah_3_5bit_score 6 (n=1, σ=0)
niah_3_5bit_total 6 (n=1, σ=0)
Parameters
framework mlx
implementation python
bits [2.5
Hypothesis

MLX TurboQuant preserves needle-in-haystack retrieval

Reference

https://github.com/Prince-Canuma

Tags
Subject
Model: Qwen3.5-35B-A3B
Baseline Comparison
niah 100% retrieval at both bit widths
Instances (1 reproduction)
apple-silicon-baselines Prince_Canuma Apple Silicon

Same-day implementation (Mar 24). 6/6 NIAH at both 2.5-bit and 3.5-bit. Zero accuracy loss confirmed. Metal kernel WIP.

niah_2_5bit_score 6 niah_2_5bit_total 6 niah_3_5bit_score 6 niah_3_5bit_total 6