AI Agent Hub
Back to models
Falcon-H1-Tiny-90M-Instruct-Curriculum logo

Falcon-H1-Tiny-90M-Instruct-Curriculum

Open Source tiiuae Released 2026-01-15
-- 0.09B params 262.1K context Open Source

About this model

drawing

Table of Contents

  1. TL;DR
  2. Model Details
  3. Training Details
  4. Usage
  5. Evaluation
  6. Citation

TL;DR

Model Details

Model Description

  • Developed by: https://www.tii.ae
  • Model type: Causal decoder-only
  • Architecture: Hybrid Transformers + Mamba architecture
  • Language(s) (NLP): English
  • Number of Parameters: 90M
  • License: Falcon-LLM License

Training details

For more details about the training protocol of this model, please refer to the Falcon-H1-Tiny technical blogpost.

Usage

Currently to use this model you can either rely on Hugging Face transformers, vLLM, sglang, llama.cpp, ollama or mlx library.

Inference

🤗 transformers

Refer to the snippet below to run H1 models using 🤗 transformers:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "tiiuae/Falcon-H1-Tiny-90M-Instruct-Curriculum"

model = AutoModelForCausalLM.from_pretrained(
  model_id,
  torch_dtype=torch.bfloat16,
  device_map="auto"
)

# Perform text generation

or

transformers serve tiiuae/Falcon-H1-Tiny-90M-Instruct-Curriculum

llama.cpp

You can find all GGUF files compatible with llama.cpp under our official collection - an example setup could be:

brew install llama.cpp 
pip install huggingface_hub 
hf download tiiuae/Falcon-H1-Tiny-90M-Instruct-Curriculum Falcon-H1-Tiny-90M-Instruct-Curriculum-Q8_0.gguf --local-dir ./ 
llama-cli ./ Falcon-H1-Tiny-90M-Instruct-Curriculum-Q8_0.gguf -cnv

ollama

ollama run hf.co/tiiuae/Falcon-H1-Tiny-90M-Instruct-Curriculum:Q8_0

Apple mlx

mlx_lm.chat --model tiiuae/Falcon-H1-Tiny-90M-Instruct-Curriculum

vLLM

For vLLM, simply start a server by executing the command below:

# pip install vllm>=0.9.0
vllm serve tiiuae/Falcon-H1-Tiny-90M-Instruct-Curriculum --tensor-parallel-size 2 --data-parallel-size 1

sglang

python -m sglang.launch_server \
  --model ttiiuae/Falcon-H1-Tiny-90M-Instruct-Curriculum \
  --tensor-parallel-size 1

Evaluation

For detailed evaluation of Falcon-H1-Tiny series, please refer to our technical blogpost

Useful links

Citation

If the Falcon-H1-Tiny family of models were helpful to your work, feel free to give us a cite.

@misc{falcon_h1_tiny,
  title={Falcon-H1-Tiny: A series of extremely small, yet powerful language models redefining capabilities at small scale},
  author={Falcon-LLM Team},
  year={2026}, 
}

Benchmark Scores

ARC
23.89
BBH
4.13
GPQA
4.51
MATH
3.6
MBPP
7.93
MMLU
24.81
GSM8K
20.24
IFEval
53.47
MMLU-Pro
1.82
HellaSwag
35.63
HumanEval
7.31
LiveBench
12.4
TruthfulQA
42.31

Technical Specs

  • Parameters: 0.09B
  • Architecture: Hybrid Transformer-Mamba
  • Context Window: 262,144 tokens
  • Input Modalities: text

Hardware Requirements

  • VRAM: 1.0 GB
  • Compute: CPU or single consumer GPU with 1GB+ VRAM (BF16 weights ~180MB)