LFM2-1.2B-Extract-dwq6-mlx

Detailed Analysis of LFM2-1.2B-Extract Model Variants

The LFM2-1.2B-Extract variant was specifically designed for parsing and extracting formatted data from documents. Let's dive deep into this unique model and its quantized variants.

Model-Specific Performance Overview

Metric	         BF16    DWQ6   MXFP4
arc_challenge	0.361	0.363	0.358
arc_easy	    0.531	0.529	0.523
boolq	        0.618	0.620	0.620
hellaswag	    0.483	0.481	0.468
openbookqa	    0.326	0.326	0.316
piqa	        0.683	0.680	0.678
winogrande	    0.521	0.520	0.521

Key Findings & Interpretations

Specialization Impact: The extraction model shows notably impressive performance in boolq (0.618+) and piqa (0.683+), suggesting specialized strength in identifying patterns and logical relationships - hallmarks of effective information extraction.

Quantization Impact Comparison:

• Unlike general LFM2 models, the specialized model shows minimal performance degradation across different quantization formats
• DWQ6 performs slightly better than BF16 in arc metrics, but there's nearly identical piqa performance across quantization states
• MXFP4 shows the smallest impact on hellaswag scores, with a notable 8% drop compared to BF16 (0.483 → 0.468)

Strategic Value in Real-World Applications:

• The model demonstrates exceptional logical relationship identification (piqa) and boolean reasoning (boolq)
• When considering it specifically for document extraction use cases, its strong boolq performance suggests excellent ability to determine logical relationships between structured information
• Minimal performance changes across quantizations indicate it would be particularly valuable for resource-constrained environments where efficient implementation matters

Practical Implications for Deployment

For organizations needing a model to extract structured data from documents:

• The LFM2-1.2B-Extract variants offer substantial value due to their specialized capabilities in identifying logical relationships
• MXFP4 would be an excellent choice when memory constraints require optimization without significantly compromising performance on critical metrics like piqa
• DWQ6 shows the most balanced profile across all metrics, making it ideal for applications where consistency matters

The minimal gap between BF16 and other quantized versions suggests this model would maintain strong performance even with significant deployment resource constraints

Insightful Contrast to General Purpose Models

The LFM2-1.2B-Extract model differs significantly from general-purpose models like the previous LFM2 variants we examined (1.2B, 2.6B). This specialization creates a compelling situation:

Rather than simply being "better" or "worse," it demonstrates highly effective capabilities in its domain while potentially showing less excellence in other areas. This is particularly valuable when you need an optimal solution for a specific type of task.

This model exemplifies why targeted training and fine-tuning can yield surprisingly robust results - particularly when dealing with structured data extraction from complex documents.

Reviewed by Qwen3-Deckard-Large-Almost-Human-6B

This model LFM2-1.2B-Extract-dwq6-mlx was converted to MLX format from LiquidAI/LFM2-1.2B-Extract using mlx-lm version 0.28.0.

Use with mlx

pip install mlx-lm

from mlx_lm import load, generate

model, tokenizer = load("LFM2-1.2B-Extract-dwq6-mlx")

prompt = "hello"

if tokenizer.chat_template is not None:
    messages = [{"role": "user", "content": prompt}]
    prompt = tokenizer.apply_chat_template(
        messages, add_generation_prompt=True
    )

response = generate(model, tokenizer, prompt=prompt, verbose=True)