File size: 7,041 Bytes
c83ff7e
80ebc50
df47971
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
80ebc50
c83ff7e
 
 
80ebc50
 
2064ed2
80ebc50
 
c83ff7e
 
2064ed2
 
80ebc50
 
 
 
 
 
 
c83ff7e
 
 
 
 
 
 
 
 
80ebc50
c83ff7e
 
 
 
 
80ebc50
 
 
 
 
c83ff7e
 
 
df47971
c83ff7e
 
 
1a21757
c83ff7e
 
 
 
 
 
 
 
 
a275fec
 
 
 
 
c83ff7e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a275fec
c83ff7e
 
 
 
 
1a21757
c83ff7e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
df47971
2cfa502
c83ff7e
 
 
 
 
 
df47971
 
 
 
 
c83ff7e
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
---
base_model: tiiuae/Falcon-H1-1.5B-Base
language:
- ar
- cs
- de
- en
- es
- fr
- hi
- it
- ja
- ko
- nl
- pl
- pt
- ro
- ru
- sv
- ur
- zh
library_name: transformers
license: other
license_name: falcon-llm-license
license_link: https://falconllm.tii.ae/falcon-terms-and-conditions.html
tags:
- falcon-h1
inference: true
pipeline_tag: text-generation
paper: tiiuae/falcon-h1
---

<img src="https://huggingface.co/datasets/tiiuae/documentation-images/resolve/main/falcon_mamba/falcon-h1-logo.png" alt="drawing" width="800"/>

**Falcon-H1** is a new series of large language models (LLMs) featuring hybrid architecture designs optimized for both high performance and efficiency across diverse use cases. Unlike earlier Falcon models built solely on Transformer or Mamba architectures, Falcon-H1 adopts a parallel hybrid approach that combines Transformer-based attention with State Space Models (SSMs), known for superior long-context memory and computational efficiency. These models excel across reasoning, mathematics, multilingual tasks, instruction following, and scientific knowledge.

*   **Paper:** [Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance](https://huggingface.co/papers/2507.22448)
*   **GitHub Repository:** [https://github.com/tiiuae/Falcon-H1](https://github.com/tiiuae/Falcon-H1)
*   **Project Documentation:** [https://tiiuae.github.io/Falcon-H1/](https://tiiuae.github.io/Falcon-H1/)

# Table of Contents

0. [TL;DR](#TL;DR)
1. [Model Details](#model-details)
2. [Training Details](#training-details)
3. [Usage](#usage)
4. [Evaluation](#evaluation)
5. [Citation](#citation)

# TL;DR
Falcon-H1 is the latest evolution in the Falcon family of large language models, built upon an advanced hybrid architecture that integrates both State Space Models (SSMs) and Attention Mechanisms in each block. These models span from 500 million to 34 billion parameters, offering high performance and efficiency. They are optimized for diverse use cases, trained with support for 18 core languages (scalable to 100+), and achieve state-of-the-art multilingual and reasoning performances in instruction following, maths, coding, and scientific knowledge tasks.

# Model Details

## Model Description

-   **Developed by:** [https://www.tii.ae](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](https://falcon-lm.github.io/blog/falcon-h1/) and [Technical Report](https://arxiv.org/abs/2507.22448).

# Usage

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

## Inference

Make sure to install the latest version of `transformers` or `vllm`, eventually install these packages from source:

```bash
pip install git+https://github.com/huggingface/transformers.git
```

For vLLM, make sure to install `vllm>=0.9.0`:

```bash
pip install "vllm>=0.9.0"
```

### 🤗 transformers

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

```python
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"
)

# Perform text generation
```

### vLLM

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

```
# pip install vllm>=0.9.0
vllm serve tiiuae/Falcon-H1-1B-Instruct --tensor-parallel-size 2 --data-parallel-size 1
```

### `llama.cpp`

You can find all GGUF files compatible with `llama.cpp` under [our official collection](https://huggingface.co/collections/tiiuae/falcon-h1-6819f2795bc406da60fab8df)

# Evaluation

Falcon-H1 series perform very well on a variety of tasks, including reasoning tasks. 

| Tasks | Falcon-H1-1.5B | Qwen3-1.7B | Qwen2.5-1.5B | Gemma3-1B | Llama3.2-1B | Falcon3-1B |
| --- | --- | --- | --- | --- | --- | --- |
| **General**  | | | | | |
| BBH | **46.47** | 35.18 | 42.41 | 35.86 | 33.21 | 34.47 |
| ARC-C | 42.06 | 34.81 | 40.53 | 34.13 | 34.64 | **43.09** |
| TruthfulQA | 45.98 | **49.39** | 47.05 | 42.17 | 42.08 | 42.31 |
| HellaSwag | **63.33** | 49.27 | 62.23 | 42.24 | 55.3 | 58.53 |
| MMLU | **62.03** | 57.04 | 59.76 | 40.87 | 45.93 | 46.1 |
| **Math**  | | | | | |
| GSM8k | **74.98** | 69.83 | 57.47 | 42.38 | 44.28 | 44.05 |
| MATH-500 | **74.0** | 73.0 | 48.4 | 45.4 | 13.2 | 19.8 |
| AMC-23 | 43.59 | **46.09** | 24.06 | 19.22 | 7.19 | 6.87 |
| AIME-24 | 11.25 | **12.5** | 2.29 | 0.42 | 1.46 | 0.41 |
| AIME-25 | **9.58** | 8.12 | 1.25 | 1.25 | 0.0 | 0.21 |
| **Science**  | | | | | |
| GPQA | 26.34 | 27.68 | 26.26 | **28.19** | 26.59 | 26.76 |
| GPQA_Diamond | **35.19** | 33.33 | 25.59 | 21.55 | 25.08 | 31.31 |
| MMLU-Pro | **37.8** | 23.54 | 28.35 | 14.46 | 16.2 | 18.49 |
| MMLU-stem | **64.13** | 54.3 | 54.04 | 35.39 | 39.16 | 39.64 |
| **Code**  | | | | | |
| HumanEval | **68.29** | 67.68 | 56.1 | 40.85 | 34.15 | 22.56 |
| HumanEval+ | **61.59** | 60.96 | 50.61 | 37.2 | 29.88 | 20.73 |
| MBPP | **64.81** | 58.73 | **64.81** | 57.67 | 33.6 | 20.63 |
| MBPP+ | **56.35** | 49.74 | 56.08 | 50.0 | 29.37 | 17.2 |
| LiveCodeBench | **17.61** | 14.87 | 12.52 | 5.09 | 2.35 | 0.78 |
| CRUXEval | **39.57** | 18.88 | 34.76 | 12.7 | 0.06 | 15.58 |
| **Instruction Following**  | | | | | |
| IFEval | **80.66** | 70.77 | 45.33 | 61.48 | 55.34 | 54.26 |
| Alpaca-Eval | **28.18** | 21.89 | 9.54 | 17.87 | 9.38 | 6.98 |
| MTBench | **8.46** | 7.61 | 7.1 | 7.03 | 6.37 | 6.03 |
| LiveBench | 34.13 | **40.73** | 21.65 | 18.79 | 14.97 | 14.1 |

You can check more in detail on our [our release blogpost](https://falcon-lm.github.io/blog/falcon-h1/), detailed benchmarks.

# Useful links

- View [our release blogpost](https://falcon-lm.github.io/blog/falcon-h1/).
- View [our technical report](https://arxiv.org/abs/2507.22448).
- Feel free to join [our discord server](https://discord.gg/trwMYP9PYm) if you have any questions or to interact with our researchers and developers.

# Citation

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

```
@article{falconh1,
    title={Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance},
    author={Jingwei Zuo and Maksim Velikanov and Ilyas Chahed and Younes Belkada and Dhia Eddine Rhayem and Guillaume Kunsch and Hakim Hacid and Hamza Yous and Brahim Farhat and Ibrahim Khadraoui and Mugariya Farooq and Giulia Campesan and Ruxandra Cojocaru and Yasser Djilali and Shi Hu and Iheb Chaabane and Puneesh Khanna and Mohamed El Amine Seddik and Ngoc Dung Huynh and Phuc Le Khac and Leen AlQadi and Billel Mokeddem and Mohamed Chami and Abdalgader Abubaker and Mikhail Lubinets and Kacper Piskorski and Slim Frikha},
    journal = {arXiv preprint arXiv:2507.22448},
    year={2025}
}
```