💧Liquid LFM2.5: How To Run & Fine-tune

Run and fine-tune LFM2.5 Instruct and Vision locally on your device!

Liquid AI releases LFM2.5, including their instruct and vision model. LFM2.5-1.2B-Instruct is a 1.17B parameter hybrid reasoning model trained on 28T tokens and RL, delivering best-in-class performance at the 1B scale for instruction following, tool use, and agentic tasks.

LFM2.5 runs on under 1GB RAM and achieves 239 tok/s decode on AMD CPU. You can also fine-tune it locally with Unsloth.

Text LFM2.5-InstructVision LFM2.5-VL

Dynamic GGUFs	16-bit Instruct
LFM2.5-1.2B-Instruct-GGUF	LFM2.5-1.2B-Instruct

Model Specifications:

• Parameters: 1.17B

• Architecture: 16 layers (10 double-gated LIV convolution blocks + 6 GQA blocks)

• Training Budget: 28T tokens

• Context Length: 32,768 tokens

• Vocabulary Size: 65,536

• Languages: English, Arabic, Chinese, French, German, Japanese, Korean, Spanish

⚙️ Usage Guide

Liquid AI recommends these settings for inference:

• temperature = 0.1

• top_k = 50

• top_p = 0.1

• repetition_penalty = 1.05

• Maximum context length: 32,768

Chat Template Format

LFM2.5 uses a ChatML-like format:

tokenizer.apply_chat_template([
    {"role": "system", "content": "You are a helpful assistant trained by Liquid AI."},
    {"role": "user", "content": "What is C. elegans?"},
], add_generation_prompt=True, tokenize=False)

LFM2.5 chat template:

<|startoftext|><|im_start|>system
You are a helpful assistant trained by Liquid AI.<|im_end|>
<|im_start|>user
What is C. elegans?<|im_end|>
<|im_start|>assistant

Tool Use

LFM2.5 supports function calling with special tokens <|tool_call_start|> and <|tool_call_end|>. Provide tools as a JSON object in the system prompt:

<|startoftext|><|im_start|>system
List of tools: [{"name": "get_weather", "description": "Gets the current weather", "parameters": {"type": "object", "properties": {"city": {"type": "string"}}}}]<|im_end|>
<|im_start|>user
What's the weather in Paris?<|im_end|>
<|im_start|>assistant
<|tool_call_start|>[get_weather(city="Paris")]<|tool_call_end|>

🖥️ Run LFM2.5-1.2B-Instruct

📖 llama.cpp Tutorial (GGUF)

1. Build llama.cpp

Obtain the latest llama.cpp from GitHub. Change -DGGML_CUDA=ON to -DGGML_CUDA=OFF if you don't have a GPU.

apt-get update
apt-get install pciutils build-essential cmake curl libcurl4-openssl-dev -y
git clone https://github.com/ggml-org/llama.cpp
cmake llama.cpp -B llama.cpp/build \
    -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON
cmake --build llama.cpp/build --config Release -j --clean-first --target llama-cli llama-server
cp llama.cpp/build/bin/llama-* llama.cpp

2. Run directly from Hugging Face

./llama.cpp/llama-cli \
    -hf LiquidAI/LFM2.5-1.2B-Instruct-GGUF:Q4_K_M \
    --jinja --ctx-size 32768 \
    --temp 0.1 --top-k 50 --top-p 0.1 --repeat-penalty 1.05

3. Or download the model first

import os
os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="LiquidAI/LFM2.5-1.2B-Instruct-GGUF",
    local_dir="LiquidAI/LFM2.5-1.2B-Instruct-GGUF",
    allow_patterns=["*Q4_K_M*"],
)

4. Run in conversation mode

./llama.cpp/llama-cli \
    --model LiquidAI/LFM2.5-1.2B-Instruct-GGUF/LFM2.5-1.2B-Instruct-Q4_K_M.gguf \
    --ctx-size 32768 \
    --n-gpu-layers 99 \
    --seed 3407 \
    --prio 2 \
    --temp 0.1 \
    --top-k 50 \
    --top-p 0.1 \
    --repeat-penalty 1.05 \
    --jinja

🦥 Fine-tuning LFM2.5 with Unsloth

Unsloth supports fine-tuning LFM2.5 models. The 1.2B model fits comfortably on a free Colab T4 GPU. Training is 2x faster with 50% less VRAM.

Free Colab Notebook:

• LFM2.5-1.2B-Instruct SFT LoRA notebook

• LFM2.5-1.2B-Instruct GRPO LoRA notebook

• LFM2.5-1.2B-Base Continued Pretraining (text completion) notebook

• LFM2.5-1.2B-Base Continued Pretraining (translation) notebook

LFM2.5 is recommended for agentic tasks, data extraction, RAG, and tool use. It is not recommended for knowledge-intensive tasks or programming.

Unsloth Config for LFM2.5

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="LiquidAI/LFM2.5-1.2B-Instruct",
    max_seq_length=4096,
    load_in_4bit=False,
)

model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    target_modules = ["q_proj", "k_proj", "v_proj", "out_proj", "in_proj",
                      "w1", "w2", "w3"],
    lora_alpha=16,
    lora_dropout=0,
    bias="none",
    use_gradient_checkpointing="unsloth",
    random_state=3407,
)

Training Setup

from trl import SFTTrainer
from transformers import TrainingArguments
from unsloth import is_bfloat16_supported

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="text",
    max_seq_length=4096,
    dataset_num_proc=2,
    packing=False,
    args=TrainingArguments(
        per_device_train_batch_size=2,
        gradient_accumulation_steps=4,
        warmup_steps=5,
        max_steps=60,
        learning_rate=2e-4,
        fp16=not is_bfloat16_supported(),
        bf16=is_bfloat16_supported(),
        logging_steps=1,
        optim="adamw_8bit",
        weight_decay=0.01,
        lr_scheduler_type="linear",
        seed=3407,
        output_dir="outputs",
    ),
)

trainer.train()

Save and Export

# Save LoRA adapters
model.save_pretrained("lfm25_lora")
tokenizer.save_pretrained("lfm25_lora")

# Merge and save to 16bit
model.save_pretrained_merged("lfm25_merged", tokenizer, save_method="merged_16bit")

# Export to GGUF
model.save_pretrained_gguf("lfm25_gguf", tokenizer, quantization_method="q4_k_m")

🎉 llama-server Serving & Deployment

To deploy LFM2.5 for production with an OpenAI-compatible API:

./llama.cpp/llama-server \
    --model LiquidAI/LFM2.5-1.2B-Instruct-GGUF/LFM2.5-1.2B-Instruct-Q4_K_M.gguf \
    --alias "LiquidAI/LFM2.5-1.2B-Instruct" \
    --threads -1 \
    --n-gpu-layers 99 \
    --ctx-size 32768 \
    --port 8001 \
    --temp 0.1 \
    --top-k 50 \
    --top-p 0.1 \
    --repeat-penalty 1.05 \
    --jinja

Test with OpenAI client:

from openai import OpenAI

client = OpenAI(
    base_url="http://127.0.0.1:8001/v1",
    api_key="sk-no-key-required",
)

completion = client.chat.completions.create(
    model="LiquidAI/LFM2.5-1.2B-Instruct",
    messages=[{"role": "user", "content": "What is 2+2?"}],
)
print(completion.choices[0].message.content)

📊 Benchmarks

LFM2.5-1.2B-Instruct delivers best-in-class performance at the 1B scale and offers fast CPU inference with low memory usage:

💧 Liquid LFM2.5-1.2B-VL Guide

LFM2.5-VL-1.6B is a vision LLM built on top of LFM2.5-1.2B-Base and tuned for stronger real-world performance. You can now fine-tune it locally with Unsloth.

Running TutorialFine-tuning Tutorial

Dynamic GGUFs	16-bit Instruct
LFM2.5-VL-1.6B-GGUF	LFM2.5-VL-1.6B

Model Specifications:

• LM Backbone: LFM2.5-1.2B-Base

• Vision encoder: SigLIP2 NaFlex shape-optimized 400M

• Context length: 32,768 tokens

• Vocabulary size: 65,536

• Languages: English, Arabic, Chinese, French, German, Japanese, Korean, and Spanish

• Native resolution processing: Handles images up to 512×512 pixels without upscaling and preserves non-standard aspect ratios without distortion

• Tiling strategy: Splits large images into non-overlapping 512×512 patches and includes thumbnail encoding for global context

• Inference-time flexibility: User-tunable maximum image tokens and tile count for speed/quality tradeoff without retraining

⚙️ Usage Guide

Liquid AI recommends these settings for inference:

• Text: temperature=0.1, min_p=0.15, repetition_penalty=1.05

• Vision: min_image_tokens=64, max_image_tokens=256, do_image_splitting=True

Chat Template Format

LFM2.5-VL uses a ChatML-like format:

tokenizer.apply_chat_template([
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": "What's in this image?"}
        ]
    },
    {"role": "assistant", "content": "I can see a cat sitting on a couch."}
], tokenize=False)

LFM2.5-VL chat template:

<|startoftext|><|im_start|>system
You are a helpful multimodal assistant by Liquid AI.<|im_end|>
<|im_start|>user
<image>Describe this image.<|im_end|>
<|im_start|>assistant
This image shows a Caenorhabditis elegans (C. elegans) nematode.<|im_end|>

🖥️ Run LFM2.5-VL-1.6B

📖 llama.cpp Tutorial (GGUF)

1. Build llama.cpp

Obtain the latest llama.cpp from GitHub. Change -DGGML_CUDA=ON to -DGGML_CUDA=OFF if you don't have a GPU.

apt-get update
apt-get install pciutils build-essential cmake curl libcurl4-openssl-dev -y
git clone https://github.com/ggml-org/llama.cpp
cmake llama.cpp -B llama.cpp/build \
    -DBUILD_SHARED_LIBS=OFF -DGGML_CUDA=ON -DLLAMA_CURL=ON
cmake --build llama.cpp/build --config Release -j --clean-first --target llama-cli llama-server
cp llama.cpp/build/bin/llama-* llama.cpp

2. Run directly from Hugging Face

./llama.cpp/llama-cli \
  -hf LiquidAI/LFM2.5-VL-1.6B-GGUF:Q4_0 \
  --image test_image.jpg \
  --image-max-tokens 64 \
  -p "What's in this image?" \
  -n 128

🦥 Fine-tuning LFM2.5-VL with Unsloth

Unsloth supports fine-tuning LFM2.5 models. The 1.6B model fits comfortably on a free Colab T4 GPU. Training is 2x faster with 50% less VRAM.

Free Colab Notebook:

• LFM2.5-VL-1.6B SFT LoRA notebook

Unsloth Config for LFM2.5

from unsloth import FastVisionModel
import torch

model, tokenizer = FastVisionModel.from_pretrained(
    model_name = "LiquidAI/LFM2.5-VL-1.6B",
    max_seq_length = 4096, 
    load_in_4bit = False, 
)

model = FastVisionModel.get_peft_model(
    model,
    finetune_vision_layers     = False, # Set to False for now
    finetune_language_layers   = True, # False if not finetuning language layers
    finetune_attention_modules = True, # False if not finetuning attention layers
    finetune_mlp_modules       = True, # False if not finetuning MLP layers
    r = 16,         
    lora_alpha = 16,
    lora_dropout = 0,
    bias = "none",
)

Training Setup

from unsloth.trainer import UnslothVisionDataCollator
from trl import SFTTrainer, SFTConfig

FastVisionModel.for_training(model) # Enable for training!

trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    data_collator = UnslothVisionDataCollator(model, tokenizer), # Must use!
    train_dataset = converted_dataset,
    args = SFTConfig(
        per_device_train_batch_size = 2,
        gradient_accumulation_steps = 4,
        warmup_steps = 5,
        max_steps = 30,# num_train_epochs = 1, # Set this instead of max_steps for full train
        learning_rate = 2e-4,
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.001,
        lr_scheduler_type = "linear",
        seed = 3407,
        output_dir = "outputs",
        report_to = "none",     # For Weights and Biases
        remove_unused_columns = False,
        dataset_text_field = "",
        dataset_kwargs = {"skip_prepare_dataset": True},
        max_length = 2048,
    ),
)

trainer.train()

Save and Export

# Save LoRA adapters
model.save_pretrained("lfm25_lora")
tokenizer.save_pretrained("lfm25_lora")

# Merge and save to 16bit
model.save_pretrained_merged("lfm25_merged", tokenizer, save_method="merged_16bit")

# Export to GGUF
model.save_pretrained_gguf("lfm25_gguf", tokenizer, quantization_method="q4_k_m")

📊 Benchmarks

LFM2.5-VL-1.6B delivers best-in-class performance:

Model	MMStar	MM-IFEval	BLINK	InfoVQA (Val)	OCRBench (v2)	RealWorldQA	MMMU (Val)	MMMB (avg)	Multilingual MMBench (avg)
LFM2.5-VL-1.6B	50.67	52.29	48.82	62.71	41.44	64.84	40.56	76.96	65.90
LFM2-VL-1.6B	49.87	46.35	44.50	58.35	35.11	65.75	39.67	72.13	60.57
InternVL3.5-1B	50.27	36.17	44.19	60.99	33.53	57.12	41.89	68.93	58.32
FastVLM-1.5B	53.13	24.99	43.29	23.92	26.61	61.56	38.78	64.84	50.89

📚 Resources

• Liquid AI Blog Post

• LFM2 Technical Report (arXiv)

• Liquid AI Documentation

• Liquid Playground