Modelpublic

OpenMed/LFM2-1.2B-Longevity-4bit-mlx

sourceHugging Faceotherupdated 4d agoView on Hugging Face

0likes26downloads

Model Card

LFM2-1.2B-Longevity - MLX 4-bit

A native MLX conversion of `LiquidAI/LFM2-1.2B-Longevity`, affine-quantized to 4-bit (group size 64) with mlx_lm for on-device use on Apple Silicon. The model is Liquid AI and Insilico Medicine's Longevity-LLM fine-tune of `LiquidAI/LFM2-1.2B`; OpenMed made and published this conversion. For the 8-bit sibling, see `OpenMed/LFM2-1.2B-Longevity-8bit-mlx`.

Family at a glance: - PyTorch source (BF16): `LiquidAI/LFM2-1.2B-Longevity` - 2.18 GiB weights - MLX 4-bit (this repo): `OpenMed/LFM2-1.2B-Longevity-4bit-mlx` - Apple Silicon, 0.61 GiB weights - MLX 8-bit: `OpenMed/LFM2-1.2B-Longevity-8bit-mlx` - Apple Silicon, 1.16 GiB weights

Why 4-bit?

	BF16 source	This repo (Q4)
`model.safetensors` size	2.18 GiB	0.61 GiB (3.6x smaller)
Bits per weight, measured from the tensors	16	4.50
Parameters	1.17 B	1.17 B (unchanged)

4-bit is the rung to reach for when memory is the constraint - a phone or a base-model Mac - and the weights above are the whole download.

Quantization

Field	Value
Bits	4
Group size	64
Mode	affine MLX weight-only quantization
Tool	`mlx_lm.convert` (mlx-lm 0.31.3), `-q --q-bits 4 --q-group-size 64`
Quantized modules	every linear projection and the tied token embedding
Kept in BF16	RMSNorm scales and the short-convolution kernels

The tokenizer, chat template and generation defaults are the upstream files, unchanged.

What it does

Longevity-LLM (L-LLM) is a family of compact, domain-adapted language models for interpreting heterogeneous aging-biology data - genomic, proteomic and clinical - developed jointly by Insilico Medicine and Liquid AI. This checkpoint was produced by full-parameter supervised fine-tuning of LiquidAI/LFM2-1.2B on aging-related multi-omics and clinical data, and accompanies the study An Open Benchmark and Language Models for AI in Aging Biology (Zhavoronkov et al., 2026). See the upstream model card for the training corpus and evaluation.

Prompts use a ChatML-style template with a dynamic-thinking switch: suffix a user turn with /think for a reasoning trace or /no_think for a direct answer. The template is bundled, so apply_chat_template and mlx_lm handle it for you.

Architecture

Field	Value
Source model type	`lfm2` (`Lfm2ForCausalLM`)
Design	Hybrid Liquid model: gated short convolutions with 6 grouped-query attention layers out of 16
Hidden size	2048
Layers	16 (6 attention, 10 convolution)
Attention	32 query heads / 8 KV heads
Feed-forward width	12,288
Context length	32,768 tokens per the upstream card (`max_position_embeddings` 128,000)
Vocabulary	65,536, tied input/output embeddings
Language	English

File set

File	Size	Purpose
`model.safetensors`	0.61 GiB	MLX weights, affine-quantized
`model.safetensors.index.json`	22.9 KiB	Tensor index
`config.json`	1.7 KiB	Model config with the `quantization` block
`modeling_lfm2.py`	35.7 KiB	Reference PyTorch modeling file carried from upstream; not used by MLX
`chat_template.jinja`	2.5 KiB	ChatML-style chat template, unchanged from upstream
`tokenizer.json`	4.5 MiB	Tokenizer, unchanged from upstream
`tokenizer_config.json`	0.5 KiB	Tokenizer metadata, unchanged from upstream
`generation_config.json`	0.2 KiB	Default sampling settings, unchanged from upstream
`LICENSE`	10.3 KiB	LFM Open License v1.0 (upstream copy)

Weights SHA-256: d4968839d8c62d8fad070aaadfb6d41eb8b7ed32d0e69e8c5096a12e64b185e2

Quick start

With mlx-lm

bash

pip install -U mlx-lm
mlx_lm.generate --model OpenMed/LFM2-1.2B-Longevity-4bit-mlx \
  --prompt "Which biomarkers in a routine blood panel are most informative about biological age, and why? /no_think" \
  --max-tokens 400

python

from mlx_lm import load, generate

model, tokenizer = load("OpenMed/LFM2-1.2B-Longevity-4bit-mlx")

messages = [
    {"role": "system", "content": "You are a biomedical AI specialized in aging biology, trained on genomic, proteomic, and clinical data."},
    {"role": "user", "content": "A 58-year-old has HbA1c 6.1%, hs-CRP 3.4 mg/L and an epigenetic age 6 years above chronological age. What do these point to, and what would you want measured next? /no_think"},
]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True)
print(generate(model, tokenizer, prompt=prompt, max_tokens=500))

With LM Studio

Download the repository into your LM Studio models directory (or search for OpenMed/LFM2-1.2B-Longevity-4bit-mlx in the app); the MLX engine loads it directly.

Outputs are model predictions for research use, not clinical advice.

Provenance and license

This repository is a quantized derivative of LiquidAI/LFM2-1.2B-Longevity. model.safetensors was produced by mlx_lm.convert from the upstream BF16 weights and config.json carries the added quantization block; every other file is an unchanged upstream copy. It is distributed under the upstream LFM Open License v1.0: free for research and for commercial use by entities below the license's revenue threshold (US$10M annual revenue), and not licensed for commercial use above it - read the LICENSE before shipping. Copyright in the model remains with Liquid AI, Inc. and Insilico Medicine; OpenMed claims none.

Citation

Please cite the upstream work:

bibtex

@article{zhavoronkov2026longevitybench,
  title   = {An Open Benchmark and Language Models for AI in Aging Biology},
  author  = {Zhavoronkov, Alex and Naumov, Vladimir and Sidorenko, Denis and Aliper, Alex and Aladinskiy, Vladimir and Hasani, Ramin and Amini, Alexander and Nasto, Katerina and Reymond, Mathieu and Shayakhmetov, Rim and Miftakhutdinov, Zulfat and Gladyshev, Vadim N. and Galkin, Fedor},
  journal = {Cell},
  volume  = {189},
  pages   = {5980--5994},
  year    = {2026},
  doi     = {10.1016/j.cell.2026.08.026},
  url     = {https://www.cell.com/cell/fulltext/S0092-8674(26)00999-2},
}