Falcon 2Highlight 1
Two open-weight checkpoints: Falcon 2 11B for text, Falcon 2 11B VLM for image-to-text.
May 2024 - Open weights
Meet Falcon 2: open-weight 11B text and vision models from TII
Falcon 2 11B is an 11-billion-parameter model published by TII under the TII Falcon License 2.0, with a vision-language sibling built on the same backbone. Both ship as raw open weights, so you host them yourself. falcon2.lol is an independent third party and every figure below links back to its source.
Multilingual Vision-to-language 11B & 11B VLM Runs on one GPU TII Falcon License 2.0
What Falcon 2 actually ships
Two checkpoints, both open weights, both runnable on one GPU. These are the three things worth knowing before you pick Falcon 2 for a project.
Highlight 2
On the Hugging Face Open LLM Leaderboard, Falcon 2 11B averaged 64.28 against 64.29 for Google's Gemma 7B, ahead of Meta's Llama 3 8B.
Highlight 3
TII has signalled more sizes and Mixture-of-Experts variants for the series; none of that is released yet.
Two checkpoints, one 11B backbone
The series starts with a text model and a vision-language model that share the same 11-billion-parameter decoder, so moving between them does not mean relearning a stack.
Falcon 2 11B
A more efficient and accessible LLM trained on 5.5 trillion tokens with 11 billion parameters.
Falcon 2 11B VLM
Distinguished by its vision-to-language model (VLM) capabilities.
Released under the TII Falcon License 2.0
Both checkpoints are published under the TII Falcon License 2.0, an Apache 2.0-based licence with an added acceptable use policy. The licence text is maintained by TII, not by us: always read the current version before shipping. More information can be found at FalconLLM.TII.ae.
Falcon 2 11B performance and positioning
TII reported that Falcon 2 11B outperformed Meta's Llama 3 8B and landed within 0.01 of Google's Gemma 7B on the Hugging Face Open LLM Leaderboard at the time of release (Falcon 2 11B 64.28 against Gemma 7B 64.29). Leaderboards move, so treat those numbers as a snapshot of May 2024 rather than a permanent ranking.
Falcon 2 11B
Average performance on the Hugging Face Leaderboard
Google Gemma 7B
First place, a difference of only 0.01 average performance
Support across languages, text, and vision inputs
Falcon 2 11B was trained mainly on English, with German, Spanish, French, Italian, Portuguese, Polish, Dutch, Romanian and Czech in the mix. TII notes that languages outside that set will not generalise well.
English, German, Spanish, French, Italian
Falcon 2 11B was trained mainly on English, with German, Spanish, French, Italian, Portuguese, Polish, Dutch, Romanian and Czech in the mix. TII notes that languages outside that set will not generalise well.
Identify and interpret images and visuals
Falcon 2 11B VLM pairs the same 11B decoder with a CLIP ViT-L/14 vision encoder and dynamic high-resolution encoding, so it can describe an image or answer questions about one. The checkpoint is English-only.
One GPU, lighter infrastructures
The 11B weights are about 22 GB in bfloat16 and roughly 7 to 8 GB at 4-bit, which is why a single 24 GB GPU is enough to serve them. That single-GPU target is the deployment story the series was designed around.
Mixture of Experts and deeper multimodal adaptation
TII has said the Falcon 2 series will be extended with Mixture-of-Experts (MoE) variants and with multimodal models in additional sizes.
MoE routes each token through a subset of specialised sub-networks instead of every parameter, which is how one system can cover many domains without paying for the full model on every token.
None of that is ours to schedule. Treat it as TII direction of travel, not a release date.
Watch this space...
Falcon 2 FAQ
Built from the official model cards, the TII Falcon License 2.0 and TII launch material. We are not TII and do not speak for them: the licence text itself is the authoritative source.
What is Falcon 2?+
Falcon 2 is an open-weight model series from TII. Falcon 2 11B is an 11-billion-parameter decoder-only language model trained on more than 5,000 billion tokens, and Falcon 2 11B VLM adds image understanding on top of the same 11B backbone.
What is the difference between Falcon 2 11B and Falcon 2 11B VLM?+
Falcon 2 11B takes text and returns text. Falcon 2 11B VLM pairs the same 11B model with a CLIP ViT-L/14 vision encoder and dynamic high-resolution image encoding, so it can also answer questions about an image. TII publishes both as raw weights rather than as a hosted API.
Can I use Falcon 2 in commercial products?+
Yes. Both checkpoints ship under the TII Falcon License 2.0, an Apache 2.0-based licence with an added acceptable use policy. Read that policy before shipping: it is the part that goes beyond plain Apache 2.0.
Does Falcon 2 support multiple languages?+
Falcon 2 11B is trained mainly on English, with additional German, Spanish, French, Italian, Portuguese, Polish, Dutch, Romanian and Czech data. The VLM is English-only. TII warns that languages outside the training mix will not generalise well.
How much hardware does Falcon 2 need?+
The 11B weights are roughly 22 GB in bfloat16, so a single 24 GB GPU runs the model, and a 4-bit quantisation fits in about 7 to 8 GB. TII positions the series for single-GPU and lighter deployments.
Is Falcon 2 11B a chat model?+
The text checkpoint on Hugging Face is a raw pretrained model and TII recommends fine-tuning it for most use cases. The VLM was built from a chat-finetuned Falcon 2 11B, so it follows instructions better out of the box. For conversation from the text base model you add the chat formatting yourself.
Where the facts on this page come from
falcon2.lol is an independent third-party site. It is not affiliated with, endorsed by, sponsored by or operated by the Technology Innovation Institute. Every figure and quotation is attributed to its original source.
Falcon 2 11B model card for parameters, training tokens, languages and licence notes.
Falcon 2 11B VLM model card for the CLIP ViT-L/14 vision encoder and the image prompt format.
FalconLLM.TII.ae for TII product pages and the TII Falcon License 2.0.
arXiv 2407.14885 for the Falcon 2 technical report.
Nothing is on sale here yet
Falcon 2 inference is not running on this site yet, so there is nothing to buy. Sign in with Google and we will email you once it actually is.