Instructions to use surajkumarmallick/Klea-1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use surajkumarmallick/Klea-1 with Transformers:

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

pipe = pipeline("text-generation", model="surajkumarmallick/Klea-1")
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("surajkumarmallick/Klea-1")
model = AutoModelForCausalLM.from_pretrained("surajkumarmallick/Klea-1", 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 surajkumarmallick/Klea-1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "surajkumarmallick/Klea-1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "surajkumarmallick/Klea-1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/surajkumarmallick/Klea-1

SGLang

How to use surajkumarmallick/Klea-1 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 "surajkumarmallick/Klea-1" \
    --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": "surajkumarmallick/Klea-1",
		"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 "surajkumarmallick/Klea-1" \
        --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": "surajkumarmallick/Klea-1",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use surajkumarmallick/Klea-1 with Docker Model Runner:
```
docker model run hf.co/surajkumarmallick/Klea-1
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

🩵 Klea-1

Lightweight in size. Heavyweight in reasoning.

✨ Overview

Klea-1 is an advanced, lightweight reasoning model developed by Suraj Kumar Mallick. It packs strong step-by-step thinking, instruction following, and long-context understanding into a footprint small enough to run on modest hardware.


Developer	Suraj Kumar Mallick
Model type	Decoder-only language model, reasoning-focused
Total parameters	543,456,768 (~540M)
Active parameters	~430M
Knowledge Cutoff	Early September 2026
Context window	256K tokens
License	Apache 2.0
Language	English

🔹 Highlights

🧠 Built to reason: produces an explicit thinking trace before its final answer.
🪶 Truly lightweight: ~540M total / ~430M active parameters.
📚 256K context: handles long documents, codebases, and extended conversations.
🎯 Instruction-tuned and preference-aligned: refined with RLHF in the final stage.
🔓 Apache 2.0: free for research and commercial use.

🏗️ Training Pipeline

Stage	Phase	What happens
1	Primary pretraining	Trained on large-scale raw web text to build broad language and world knowledge.
2	Secondary pretraining	Trained on highly curated, specialized data to sharpen domain depth and quality.
3	Reasoning + instruction tuning	Fine-tuned on reasoning data with instruction following.
4	Preference tuning	Aligned with human preferences using RLHF.

📊 Benchmarks

Reasoning & Knowledge

Benchmark	Setting	Score
GPQA Diamond	–	89.4
HLE	–	53.5
AGIEval (EM)	3–5-shot	83.5
C-Eval (EM)	5-shot	92.1
HellaSwag (EM)	0-shot	85.0
SimpleQA-Verified (EM)	25-shot	0.1

Math & Code

Benchmark	Setting	Score
GSM8K	8-shot	93.5
HumanEval (Pass@1)	0-shot	62.0

Agentic & Terminal

Benchmark	Score
Terminal-Bench 2.1	76.5
Terminal-Bench 3.0	44.0
Terminal-Bench 4.0	58.6
CyberGym	65.5

Long Context

Benchmark	Setting	Score
LongBench-V2 (EM)	1-shot	44.7

🚀 Quickstart

Install

pip install -U transformers accelerate torch

Chat with Klea-1 (reasoning mode)

Klea-1 thinks first, then answers. Its output wraps the reasoning in <think>…</think> and the final response in <answer>…</answer>.

from transformers import AutoModelForCausalLM, AutoTokenizer
import os
import re
import torch

model_name_or_path = "surajkumarmallick/Klea-1"

tokenizer = AutoTokenizer.from_pretrained(model_name_or_path)
model = AutoModelForCausalLM.from_pretrained(
    model_name_or_path, 
    device_map="auto", 
    torch_dtype=torch.bfloat16
)

messages = [
    {"role": "user", "content": "Write a short summary of the benefits of regular exercise"},
]

tokenized_chat = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,  # Ensure it returns a dictionary
    enable_thinking=True  # Toggle thinking mode (default: True)
)

# Move input dictionary keys to model device
inputs = {k: v.to(model.device) for k, v in tokenized_chat.items()}

outputs = model.generate(**inputs, max_new_tokens=2048)

output_text = tokenizer.decode(outputs[0])
print("output_text =", output_text)

think_pattern = r'<think>(.*?)</think>'
think_matches = re.findall(think_pattern, output_text, re.DOTALL)

answer_pattern = r'<answer>(.*?)</answer>'
answer_matches = re.findall(answer_pattern, output_text, re.DOTALL)

think_content = [match.strip() for match in think_matches][0] if think_matches else "No thinking process found."
answer_content = [match.strip() for match in answer_matches][0] if answer_matches else output_text
print(f"thinking_content: {think_content}\n\n")
print(f"answer_content: {answer_content}\n\n")

Thinking vs. fast mode

Mode	How to enable
Slow thinking (default)	`enable_thinking=True`, or prefix the prompt with `/think`
Fast response	`enable_thinking=False`, or prefix the prompt with `/no_think`

Recommended sampling parameters

{
  "top_k": 20,
  "top_p": 0.8,
  "repetition_penalty": 1.05,
  "temperature": 0.7
}

💬 System Prompt

For best results, use the default system prompt:

You are Klea 1, an advanced lightweight AI model developed by Suraj Kumar Mallick. Your primary goal is to assist users with their tasks accurately, clearly, and helpfully.

🎯 Intended Use

Step-by-step reasoning, math, and logic problems
Coding assistance and terminal / agentic workflows
Long-document question answering and summarization (up to 256K tokens)
On-device and low-resource deployments where larger models are impractical

⚠️ Limitations

As a compact model, factual recall on obscure or long-tail knowledge is limited (see SimpleQA-Verified). Pair it with retrieval for fact-heavy tasks.
Like all language models, Klea-1 can produce incorrect or biased outputs. Verify important results, especially in high-stakes settings.
Benchmark scores depend on prompt format, sampling settings, and evaluation harness, so results may vary in your setup.

📄 License

Klea-1 is released under the Apache License 2.0.

📖 Citation

@misc{klea1,
  title        = {Klea-1: A Lightweight Reasoning Model},
  author       = {Suraj Kumar Mallick},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/surajkumarmallick/Klea-1}}
}

🩵 Built by Suraj Kumar Mallick 🩵

Downloads last month: 362

Safetensors

Model size

0.5B params

Tensor type

BF16

Model tree for surajkumarmallick/Klea-1

Quantizations

1 model

Collection including surajkumarmallick/Klea-1

Klea

Collection

Klea — Lightweight, frontier-class AI models. • 1 item • Updated 1 day ago • 1