mlx-community/LFM2.5-8B-A1B-OptiQ-4bit

Built with mlx-optiq, the MLX-native toolkit to quantize, fine-tune, and serve LLMs locally on Apple Silicon, no PyTorch and no cloud. All OptiQ quants · Docs · LFM2.5 family

An OptiQ mixed-precision quant of LiquidAI/LFM2.5-8B-A1B, the sparse mixture-of-experts member of the LFM2.5 family. 5.46 GB on disk, down from 15.8 GB at bf16, with a 128k context.

LFM2.5-8B-A1B pairs the family's hybrid backbone — convolutional blocks interleaved with full attention, so only a few blocks carry a KV cache — with 32 routed experts and 4 active per token. OptiQ measures each layer's sensitivity and assigns per-layer bit-widths across both.

What it is

Property	Value
Base	LiquidAI/LFM2.5-8B-A1B
Architecture	lfm2_moe — hybrid conv + attention, 24 layers, 32 experts, 4 active
Method	OptiQ mixed-precision, sensitivity-driven (uniform-4-bit reference)
On disk	5.46 GB (bf16: 15.8 GB)
Context	128k

The routed experts carry most of the parameters and mostly tolerate 4-bit: 60 of the 66 fused expert projections land there, with 6 kept at 8-bit alongside attention, the router and the layer edges.

Capability Score

Six-metric mean, the standard OptiQ eval.

Metric	Score
MMLU (5-shot, 969 samples)	46.1%
GSM8K (1000 samples)	54.6%
IFEval (full set, strict)	32.9%
BFCL-V3 simple (200 calls)	24.5%
HumanEval (164 problems, pass@1)	14.0%
HashHop (long-context retrieval)	0.0%
Capability Score (mean of 6)	28.69

Grade-school maths is the strongest result here and the best in the family. Coding, world knowledge and long-context multi-hop retrieval are the weak spots.

Run it

pip install mlx-optiq
optiq serve --model mlx-community/LFM2.5-8B-A1B-OptiQ-4bit

That gives you an OpenAI and Anthropic compatible endpoint with mixed-precision KV cache, tool-call healing and prompt caching.

Liquid's recommended sampling for LFM2.5 is temperature 0.1, top_k 50, repetition_penalty 1.1. Those ship in generation_config.json in this repo and optiq serve applies them automatically, so you get the publisher's settings without passing any flags.

Or from Python:

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/LFM2.5-8B-A1B-OptiQ-4bit")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Extract every date from this email."}],
    add_generation_prompt=True, tokenize=False,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))

Links

• Project website: mlx-optiq.com
• All OptiQ quants: mlx-optiq.com/models
• PyPI: pypi.org/project/mlx-optiq
• Base model: LiquidAI/LFM2.5-8B-A1B