Liquid AI's LFM2 Beats GPT-5 and Claude on Aging Research Tasks

Liquid AI and Insilico Medicine have released two compact language models and a 17-task benchmark for aging research. Their reported results show that a domain-tuned model with 1.2 billion or 2.6 billion parameters can outperform much larger general-purpose systems on selected clinical and omics tasks. The approach could give research teams one locally deployable model interface across data types that usually require separate pipelines.

One benchmark spans five biological layers

LongevityBench contains 25,457 prompts built from public datasets across clinical measurements, DNA methylation, gene expression, plasma proteins, and genetic evidence. Its 17 tasks test whether a model can extract age-related signals from structured biological records.

The suite converts each biological record into structured text, allowing the same language model to process measurements from every domain. It uses four output formats:

• Binary classification: choose between two labels.
• Pairwise comparison: compare two records, such as identifying the older participant.
• Multiclass classification: select an age group or another category.
• Numeric regression: predict a continuous value.

Using multiple formats for the same source data helps isolate the effect of the requested output shape while preserving the underlying biology. A shared text representation simplifies the model interface, although upstream pipelines must still normalize units, encode missing values, select features, and serialize records consistently.

Small models lead several tasks

Liquid published LFM2-2.6B-Longevity and LFM2-1.2B-Longevity on Hugging Face under the LFM1.0 license. Both checkpoints are full-parameter supervised fine-tunes, meaning training updated every model weight rather than adding a small adapter.

The underlying LFM2 architecture combines multiplicative gates with short convolutions and supports a 32,768-token context window. That window can hold large serialized records, though longer genomic or omics profiles will require feature selection, chunking, or another compression strategy.

The comparison set included 18 frontier models, among them Gemini-3.1-Pro, GPT-5, and Claude Opus. These rankings describe ordering under LongevityBench’s prompts, preprocessing, and scoring rules. Task-level scores, score margins, repeated runs, and evaluation on a team’s own data remain necessary for model selection.

Three training runs become one checkpoint

For each parameter size, the teams trained three full-parameter variants for three epochs with a 32,768-token context window:

1. Longevity data only.
2. Longevity data mixed with 10% general chat data.
3. Longevity data mixed with 20% general chat data.

Equal-weight linear merging then combined the three variants within each model size. The general chat mixtures were intended to retain conversational behavior and limit overfitting to the benchmark’s rigid prompt structure. Reported results came from single-pass, zero-shot evaluation with the model’s thinking mode disabled.

Ablation tests examined whether predictions depended on biologically relevant inputs. The authors removed one defined feature group from a prompt, retained the remaining record, and measured the change in the gap between the model’s raw scores for correct and incorrect answers. Removing selected groups reduced confidence and sometimes changed the prediction. Dependence also varied across tasks, weakening the explanation that one shared shortcut drove every result. These ablations measure input sensitivity and leave causal biological understanding untested.

Local inference fits sensitive workflows

Local deployment can keep clinical and genomic records inside an organization’s infrastructure when policy or regulation restricts external API access. It also gives teams control over model versions, logging, retention, and network boundaries. Production use still requires access controls, encryption, audit trails, prompt handling rules, and task-specific validation.

At bfloat16 precision, the raw weights require roughly 2.4 GB for the 1.2B model and 5.2 GB for the 2.6B model. Actual inference consumes additional memory for the key-value cache, activations, and framework overhead, especially near the full 32,768-token context. Quantization can reduce the weight footprint at the cost of another accuracy variable to test.

A ChatML-style template controls the model’s dynamic thinking mode by appending /think or /no_think to a user turn. The following Transformers example disables thinking and decodes only the generated response:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "LiquidAI/LFM2-2.6B-Longevity"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    dtype=torch.bfloat16,
)
model.eval()

messages = [
    {
        "role": "system",
        "content": "You are a biomedical AI specialized in aging biology.",
    },
    {
        "role": "user",
        "content": "What are the hallmarks of aging? /no_think",
    },
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=500,
        do_sample=False,
        pad_token_id=tokenizer.eos_token_id,
    )

new_tokens = output[0, inputs["input_ids"].shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

The example assumes hardware with bfloat16 support; other devices may require float16, float32, or a quantized checkpoint. Supported local runtimes include vLLM, SGLang, llama.cpp, MLX, and LM Studio. The public LongevityBench dataset allows teams to reproduce the evaluation, add tasks, or test alternative models.

One model, bounded evidence

Across LongevityBench, supervised domain adaptation gave compact models an advantage over far larger general-purpose systems on several structured biomedical tasks. The results support a shared text-based model layer for clinical, epigenetic, transcriptomic, proteomic, and genetic inputs, while leaving normalization and modality-specific preprocessing in place.

Clinical deployment requires evidence beyond these leaderboard results. Prospective validation must address reliability, calibration, fairness, latency, cost, subgroup performance, and failure behavior in realistic workflows. Diagnostic or treatment use would also require appropriate clinical governance and regulatory review.

Coverage remains concentrated in the modalities and prompt patterns represented during training. Broader benchmarks, external datasets, contamination controls, and comparisons with specialized biological models will determine how well the approach generalizes. For current development work, the release provides two reproducible baselines and a public test suite for compact longevity and omics models.