Spark X2.6 4B for mlx-serve (mixed 4/8-bit)
An unofficial MLX affine mixed-precision quantization of XHToken/Spark-X2.6-4B, under its Apache-2.0 license. It requires mlx-serve v26.10.2 or later.
Quantization
Affine group-size-64 quantization: dense MLP projections are 4-bit; attention projections, token embeddings, and LM head are 8-bit. Norms, scales, and biases remain BF16. Pack size: 2.8 GiB.
Setup
mlx-serve serve --model pgmharikrishnan/Spark-X2.6-4B-MLX-Serve-mixed-4-8bit
Standard suite
| Pack | MMLU-Pro-400 | GSM8K-50 | Coding-6 | Decode | Prefill |
|---|---|---|---|---|---|
| 0x36 4-bit | 59.50% | 86.00% | 6/6 | 146.6 tok/s | 858 tok/s |
| This mixed 4/8-bit pack | 65.25% | 90.00% | 6/6 | 122.6 tok/s | 968 tok/s |
MMLU-Pro answers hit the 8,192-token cap in 61 mixed-pack cases versus 83 for the 4-bit pack: Spark uses high-effort thinking by default. Teacher-forced frozen-corpus fidelity is 85.94% top-1 agreement (mean KL 0.07757), versus 82.03% (0.15323) for the 4-bit baseline.
Coding-std
Full official coding evaluation (157-problem LiveCodeBench release_v6 window and 164 HumanEval+ problems; temperature 0, 16,384-token cap):
| Pack | LiveCodeBench pass@1 | HumanEval+ base pass@1 | HumanEval+ plus pass@1 |
|---|---|---|---|
| 0x36 4-bit | 31.21% | 72.56% | 67.68% |
| This mixed 4/8-bit pack | 36.31% | 77.44% | 74.39% |
Attribution
Spark X2.6 was created by XHToken. This is an independent MLX conversion for mlx-serve.
Disclaimer
This is an unofficial conversion. Validate it for your workload before production use.