Falcon model series: every generation compared
Falcon is the Technology Innovation Institute's open LLM line. Compare Falcon, Falcon3, Falcon-H1 and Falcon-Mamba by size, context and license.
What Falcon is and who makes it
Falcon is the open language model line from the Technology Innovation Institute (TII) in Abu Dhabi. It has changed shape three times: a first generation of dense decoder-only models trained mostly on web data, a Falcon3 generation of small and mid-size dense models, and since 2025 hybrid models that mix attention with state space layers. The license also changed: the first generation is Apache 2.0, while the later cards we opened use TII's own Falcon licenses, so read the license before shipping a derivative.
The generations in order
Falcon (2023)
Falcon launched with Falcon-7B and Falcon-40B, announced on June 5, 2023, with Falcon-180B following in September 2023. The cards for 7B and 40B list Apache 2.0 and a 2,048-token context. They are causal decoder-only models with multiquery attention, trained mainly on the RefinedWeb dataset: the 7B model saw 1,500 billion tokens and the 40B model 1,000 billion. The 7B card lists English and French as the main languages. These are base models for research or fine-tuning, and the context is short by today's standards.
Falcon3 (December 2024)
Falcon3 was released on December 17, 2024 in 1B, 3B, 7B and 10B sizes, plus a 7B Mamba variant. The announcement describes decoder-only transformers with SwiGLU, a 131K vocabulary and up to 32K context (8K for the 1B). The 7B model was pretrained on 14 trillion tokens, and the 10B was depth-upscaled with a further 2 trillion tokens. The Falcon3-7B-Instruct card lists the TII Falcon-LLM License 2.0 and English, French, Spanish and Portuguese.
Side lines
Falcon-H1 (2025)
Falcon-H1 is a hybrid of transformer attention and Mamba layers. The family on the 7B card spans 0.5B, 1.5B, 1.5B-Deep, 3B, 7B and 34B, with base and instruct versions, 18 languages, and the Falcon-LLM License. The paper behind it is dated July 2025. Later additions on the same architecture include Falcon-H1R-7B, a reasoning model trained with supervised fine-tuning on long reasoning traces and then reinforcement learning, which emits its reasoning inside think tags. Its card lists a default model length of 262,144 tokens.
Falcon-Mamba (August 2024)
Falcon-Mamba is a 7B state space model released on August 12, 2024, trained on about 5.5 trillion tokens under the TII Falcon Mamba 7B License 1.0. Because it carries a fixed-size state instead of a growing attention cache, TII reports it can generate arbitrarily long sequences with constant memory and constant time per token, and showed generation up to 130,000 tokens on one H100. Throughput and quality claims here are TII's own.
Two practical notes on the Falcon3 announcement: the 7B and 10B models target general use with web, code and STEM data, and the Mamba variant (Falcon3-Mamba-7B-Base) received an additional 1.5 trillion training tokens aimed at reasoning and math, with 32K context. The first-generation 7B card also states a minimum of 16GB of memory is recommended for inference.
Which generation to use today
- Smallest models or edge use: Falcon3 at 1B or 3B, or the smaller Falcon-H1 sizes.
- General instruction following up to 10B: Falcon3-7B or 10B, or Falcon-H1 at 7B if you want 18 languages and the hybrid design.
- Largest in the line: Falcon-H1-34B is the biggest size sourced here; Falcon-180B is the largest of the first generation.
- Reasoning: Falcon-H1R-7B, with the long default length above.
- Very long generation with flat memory: Falcon-Mamba or Falcon3-Mamba (32K context per the announcement), if your inference engine supports state space models.
- Permissive license: only the original Falcon models are Apache 2.0 in the cards we checked. Everything later carries a TII license with its own terms.
The original Falcon models are mainly useful as reference points now. Inference-engine support for hybrid and state space models varies, so confirm your engine supports the architecture before planning around it.
Running it
You can run these on Aquanode GPUs. Each generation hub lists the available sizes and memory estimates.
Falcon generations
Every Falcon generation we track, oldest first. Each hub lists all of its models with VRAM at native, FP8 and INT4 precision.
| Hub | Models | Sizes | Smallest native VRAM | Cheapest live fit for it | Est. $/hr |
|---|---|---|---|---|---|
| Falcon 1 | 4 | 6.9B to 41.8B | 16.1 GB | RTX A5000 | $0.176/hr |
| Falcon3 | 4 | 1.7B to 7.5B | 3.7 GB | RTX 3070 | $0.088/hr |
Smallest native VRAM is the lowest requirement among the publisher's own checkpoints in the hub, at the precision they are published in; see the methodology.
Other Falcon lines
Side lines from TII: specialised or companion models outside the main Falcon generations.
| Hub | Models | Sizes | Smallest native VRAM | Cheapest live fit for it | Est. $/hr |
|---|---|---|---|---|---|
| Falcon-H1 | 4 | 521M to 7.6B | 1.2 GB | RTX 3070 | $0.088/hr |
| Falcon Mamba | 2 | 7.3B | 16.3 GB | RTX A5000 | $0.176/hr |
Run and fine-tune Falcon
Per-generation guides: engine support, chat template and context flags, then fine-tune memory for LoRA and QLoRA.
- Falcon3: how to run, fine-tune
Best models by task
Falcon models appear on these ranked task pages.
Sources
- https://huggingface.co/blog/falcon
- https://huggingface.co/tiiuae/falcon-7b
- https://huggingface.co/tiiuae/falcon-40b
- https://huggingface.co/blog/falcon3
- https://huggingface.co/tiiuae/Falcon3-7B-Instruct
- https://huggingface.co/tiiuae/Falcon-H1-7B-Instruct
- https://huggingface.co/tiiuae/Falcon-H1R-7B
- https://huggingface.co/blog/falconmamba
Facts in the text above were read from these pages and are the publisher's own statements, not benchmarks run by Aquanode. Last reviewed 2026-10-07.