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/LFM2.5-2.6B. 1.93 GB on disk, down from 5.2 GB at bf16, with a 128k context.

LFM2.5 is a hybrid architecture: convolutional blocks interleaved with full attention, so only a few blocks carry a KV cache at all. OptiQ measures each layer's sensitivity and assigns per-layer bit-widths, keeping 8-bit where it matters and 4-bit elsewhere.

What it is

Property	Value
Base	LiquidAI/LFM2.5-2.6B
Architecture	`lfm2` — hybrid conv + full attention, 30 layers
Method	OptiQ mixed-precision, sensitivity-driven (bf16 reference)
On disk	1.93 GB (bf16: 5.2 GB)
Context	128k

Capability Score

Six-metric mean, the standard OptiQ eval.

Metric	Score
MMLU (5-shot, 969 samples)	59.5%
GSM8K (1000 samples)	38.1%
IFEval (full set, strict)	33.1%
BFCL-V3 simple (200 calls)	48.5%
HumanEval (164 problems, pass@1)	18.9%
HashHop (long-context retrieval)	13.0%
Capability Score (mean of 6)	35.19

Read that profile against what the model is for. Liquid position LFM2.5-2.6B for agentic workloads, tool use, data extraction and RAG, and state plainly that it is not recommended for agentic coding or knowledge-heavy tasks. Tool calling is the strongest result here and coding the weakest, which is the shape you would expect. World knowledge at 59.5% is the highest number in the table.

Run it

bash

pip install mlx-optiq
optiq serve --model mlx-community/LFM2.5-2.6B-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 this model 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 straight from Python:

python

from mlx_lm import load, generate

model, tokenizer = load("mlx-community/LFM2.5-2.6B-OptiQ-4bit")
prompt = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Extract the dates from this invoice."}],
    add_generation_prompt=True, tokenize=False,
)
print(generate(model, tokenizer, prompt=prompt, max_tokens=512))

mlx-community/LFM2.5-2.6B-OptiQ-4bit

mlx-community/LFM2.5-2.6B-OptiQ-4bit

What it is

Capability Score

Run it

Links

mlx-community/
LFM2.5-2.6B-OptiQ-4bit