lfm2.5-1.2b-jp-mxfp4-mlx

MLX quantization of LiquidAI/LFM2.5-1.2B-JP for Apple Silicon.

Variant: Block float MX FP4
Disk size: 598 MB
Quantized by: sahilchachra

Benchmark results

Evaluated on Apple M5 Pro with MLX. Model loaded once; performance and quality measured in a single pass.

Performance

	This model	FP16 baseline
Decode tok/s (avg, long traces)	633.26	203.66
Peak memory (GB)	1.126	2.732
Disk size (MB)	598	2237

Quality

Benchmark	This model	FP16 baseline	n
JMMLU (Japanese MMLU, accuracy)	33.3%	38.3%	60
MGSM-ja (Japanese math, accuracy)	20.0%	33.3%	30
JCommonsenseQA (Japanese commonsense, accuracy)	75.0%	77.5%	40
Long-context JA (length-met rate)	100.0%	100.0%	8

Context scaling (decode tok/s)

Context length	Decode tok/s
~128 tokens	371.9
~256 tokens	364.4
~512 tokens	364.6
~1024 tokens	361.0

Usage

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("sahilchachra/lfm2.5-1.2b-jp-mxfp4-mlx")
response = generate(model, tokenizer, prompt="Your prompt here", max_tokens=256, verbose=True)

All variants in this collection

Model	Variant
sahilchachra/lfm2.5-1.2b-jp-mxfp4-mlx	Block float MX FP4 ← this model
sahilchachra/lfm2.5-1.2b-jp-optiq-5bpw-mlx	OptiQ mixed-precision (target 5.0 bpw)

Notes

• Requires Apple Silicon (M1 or later) with MLX
• Benchmarks run on Apple M5 Pro, 24 GB unified memory
• License: see LiquidAI/LFM2.5-1.2B-JP for the original model's license

Original model

See LiquidAI/LFM2.5-1.2B-JP for full model details and intended use.