Instructions to use RedBullet12/LFM3-1.7B-Kazakh with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use RedBullet12/LFM3-1.7B-Kazakh with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="RedBullet12/LFM3-1.7B-Kazakh", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("RedBullet12/LFM3-1.7B-Kazakh", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("RedBullet12/LFM3-1.7B-Kazakh", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use RedBullet12/LFM3-1.7B-Kazakh with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "RedBullet12/LFM3-1.7B-Kazakh"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "RedBullet12/LFM3-1.7B-Kazakh",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/RedBullet12/LFM3-1.7B-Kazakh

SGLang

How to use RedBullet12/LFM3-1.7B-Kazakh with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "RedBullet12/LFM3-1.7B-Kazakh" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "RedBullet12/LFM3-1.7B-Kazakh",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "RedBullet12/LFM3-1.7B-Kazakh" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "RedBullet12/LFM3-1.7B-Kazakh",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use RedBullet12/LFM3-1.7B-Kazakh with Docker Model Runner:
```
docker model run hf.co/RedBullet12/LFM3-1.7B-Kazakh
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

LFM3-1.7B-Kazakh

LiquidAI/LFM3-1.7B fine-tuned to understand and answer in Kazakh, a language the base model does not support.

Given a Kazakh prompt, the base model reasons in English, misidentifies the language as Ukrainian or Kyrgyz, and usually never answers. This finetune answers directly in Kazakh.

Setting	Value
Data	AmanMussa/kazakh-instruction-v2 (52,123 train rows, 50 held out)
Method	LoRA SFT with TRL, merged into full weights
LoRA rank / alpha	16 / 32, all linear layers
Epochs / LR	1 / 2e-4
Hardware	1x A100 (Colab), about 1 hour

Usage

from transformers import pipeline

pipe = pipeline("text-generation", model="RedBullet12/LFM3-1.7B-Kazakh", trust_remote_code=True)
out = pipe([{"role": "user", "content": "Қазақстанның астанасы қай қала?"}],
           do_sample=True, temperature=0.3, top_p=0.9, repetition_penalty=1.15,
           no_repeat_ngram_size=4, max_new_tokens=512)
print(out[0]["generated_text"][-1]["content"].replace("</think>", "").strip())

Use these generation settings. With weaker repetition control the model can loop.

Limitations

Fluent Kazakh, but weak at math and multi-step reasoning: fine-tuning on answer-only data removed the base model's thinking step.
Can fall into repetition loops without the settings above.
Training data is machine-translated Alpaca with manual fixes, so phrasing can read as translated, and some answers are factually wrong.
v1. Next: continued pretraining on native Kazakh text before SFT.
Do not use for factual questions. v1 gives fluent Kazakh but frequently wrong facts (in testing it named Beijing as Kazakhstan's capital). Fine-tuning overwrote knowledge the base model had. Use it with retrieval (RAG) for facts, or wait for v2. Released under the LFM Open License v1.0, inherited from the base model.

Downloads last month: 512

Safetensors

Model size

2B params

Tensor type

BF16

Model tree for RedBullet12/LFM3-1.7B-Kazakh

Base model

LiquidAI/LFM3-1.7B-Base

Finetuned

LiquidAI/LFM3-1.7B

Finetuned

(56)

this model

Quantizations

1 model

RedBullet12
/

LFM3-1.7B-Kazakh

LFM3-1.7B-Kazakh

Usage

Limitations

Model tree for RedBullet12/LFM3-1.7B-Kazakh

Dataset used to train RedBullet12/LFM3-1.7B-Kazakh