đ Falcon-H1 Model
Falcon-H1 is a family of hybrid-head language models developed by TII, offering high efficiency and performance across various NLP tasks.
đ Quick Start
To use the Falcon-H1 model, you can rely on Hugging Face transformers
, vLLM
, or a custom fork of the llama.cpp
library.
Installation
Make sure to install the latest version of transformers
or vLLM
. You can install these packages from source:
pip install git+https://github.com/huggingface/transformers.git
Refer to the official vLLM documentation for more details on building vLLM from source.
Inference
đ¤ transformers
Refer to the snippet below to run H1 models using đ¤ transformers:
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "tiiuae/Falcon-H1-1B-Base"
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto"
)
vLLM
For vLLM, simply start a server by executing the command below:
# pip install vllm
vllm serve tiiuae/Falcon-H1-1B-Instruct --tensor-parallel-size 2 --data-parallel-size 1
llama.cpp
While the team is working on integrating the architecture directly into the llama.cpp
library, you can install the custom fork of the library and use it directly: https://github.com/tiiuae/llama.cpp-Falcon-H1. Use the same installing guidelines as llama.cpp
.
⨠Features
- Model Type: Causal decoder-only
- Architecture: Hybrid Transformers + Mamba architecture
- Language(s) (NLP): English, Multilingual
- License: Falcon-LLM License
đĻ Installation
For installation details, refer to the "Quick Start" section above.
đģ Usage Examples
Basic Usage
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "tiiuae/Falcon-H1-1B-Base"
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
device_map="auto"
)
đ Documentation
Model Details
- Developed by: https://www.tii.ae
- Model type: Causal decoder-only
- Architecture: Hybrid Transformers + Mamba architecture
- Language(s) (NLP): English, Multilingual
- License: Falcon-LLM License
Training Details
For more details about the training protocol of this model, please refer to the Falcon-H1 technical blogpost.
Evaluation
Falcon-H1 series perform very well on a variety of tasks, including reasoning tasks.
Tasks |
Falcon-H1-7B |
Qwen3-8B |
Qwen2.5-7B |
Gemma3-12B |
Llama3.1-8B |
Falcon3-7B |
Falcon3-10B |
General |
|
|
|
|
|
|
|
BBH |
62.28 |
47.47 |
53.76 |
63.36 |
48.58 |
52.12 |
58.09 |
ARC-C |
59.98 |
42.06 |
41.38 |
51.96 |
52.39 |
54.35 |
54.44 |
TruthfulQA |
59.91 |
53.19 |
62.41 |
61.02 |
52.99 |
55.58 |
55.05 |
HellaSwag |
75.92 |
60.56 |
63.4 |
55.63 |
71.28 |
71.81 |
75.57 |
MMLU |
76.83 |
71.56 |
73.64 |
72.5 |
68.67 |
70.81 |
74.01 |
Math |
|
|
|
|
|
|
|
GSM8k |
81.65 |
78.92 |
71.95 |
87.49 |
82.49 |
81.05 |
85.06 |
MATH-500 |
73.4 |
83.8 |
75.8 |
86.2 |
45.8 |
69.0 |
68.6 |
AMC-23 |
56.72 |
70.78 |
53.91 |
66.88 |
22.81 |
40.0 |
45.78 |
AIME-24 |
16.04 |
28.33 |
12.29 |
22.5 |
5.42 |
8.75 |
9.79 |
AIME-25 |
13.96 |
19.17 |
9.58 |
18.75 |
0.42 |
6.25 |
5.42 |
Science |
|
|
|
|
|
|
|
GPQA |
36.33 |
25.84 |
31.79 |
33.98 |
32.72 |
31.21 |
33.39 |
GPQA_Diamond |
56.9 |
43.1 |
33.0 |
37.71 |
31.31 |
37.21 |
34.68 |
MMLU-Pro |
51.75 |
34.64 |
43.23 |
39.88 |
36.42 |
40.73 |
44.05 |
MMLU-stem |
77.61 |
66.89 |
69.36 |
66.54 |
59.31 |
67.43 |
70.57 |
Code |
|
|
|
|
|
|
|
HumanEval |
86.59 |
84.75 |
82.32 |
84.76 |
68.29 |
71.95 |
82.32 |
HumanEval+ |
81.1 |
79.27 |
73.78 |
75.61 |
61.59 |
65.85 |
75.0 |
MBPP |
80.69 |
71.96 |
79.63 |
85.71 |
68.25 |
77.25 |
73.28 |
MBPP+ |
68.78 |
62.7 |
68.25 |
72.22 |
55.03 |
65.87 |
64.02 |
LiveCodeBench |
35.03 |
45.6 |
32.68 |
30.92 |
15.85 |
12.72 |
19.77 |
CRUXEval |
66.51 |
72.7 |
56.9 |
67.67 |
21.57 |
55.0 |
59.57 |
Instruction Following |
|
|
|
|
|
|
|
IFEval |
85.35 |
83.43 |
75.25 |
81.51 |
77.04 |
76.59 |
78.84 |
Alpaca-Eval |
40.23 |
46.13 |
29.48 |
43.55 |
25.48 |
27.56 |
24.31 |
MTBench |
8.85 |
8.74 |
8.45 |
8.69 |
8.29 |
8.73 |
8.46 |
LiveBench |
45.74 |
56.19 |
37.13 |
49.23 |
31.73 |
32.35 |
34.3 |
You can check more in detail on our our release blogpost, detailed benchmarks.
đ License
This model is licensed under the Falcon-LLM License.
đ Citation
If the Falcon-H1 family of models were helpful to your work, feel free to give us a cite.
@misc{tiifalconh1,
title = {Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance},
url = {https://falcon-lm.github.io/blog/falcon-h1},
author = {Falcon-LLM Team},
month = {May},
year = {2025}
}
đ Useful Links