Back to Gap Map Model components / Base / pretrained models

Falcon 3

Technology Innovation Institute
open weights / Overall score: 1.7

TII's open-weights model line from Abu Dhabi. The Falcon 3 family spans 1B to 10B and was trained on 14 trillion tokens of web, code and STEM data. It has since been extended by Falcon-H1, a hybrid Mamba-transformer architecture from 0.5B to 34B with native support for eighteen languages, and by an Arabic-specialized Falcon-H1 line at 3B, 7B and 34B.

The entry is scored on the Falcon 3 generation, now dated relative to the newer Falcon-H1 family.

Openness

3 high confidence
3.0
weights
open
data
closed
code
closed
license
TII-Falcon-License-2.0(Apache-based+AUP,non-OSI)

Falcon's weights are downloadable and free to use commercially at any scale under the TII Falcon License 2.0, an Apache-based license with an acceptable-use policy rather than a usage cap. TII does not release the training corpus or training code, and the acceptable-use clause keeps the license outside the OSI definition, so the release counts as open weights rather than open source.

  • https://falconllm.tii.ae/falcon-terms-and-conditions.html recorded 2026-09-29

    TII terms page, the December 2024 TII Falcon License that the Falcon 3 cards link: "This license is, in part, based on the Apache License Version 2.0"; "Your copyright license to use the Work shall be royalty free and without charge"; "your use of the Work or any Derivative Work must comply with the Acceptable Use Policy at all times"; derivatives must state "is built using Falcon LLM technology from the Technology Innovation Institute". Searched this license for monthly, revenue caps and non-commercial terms - none; the hosting-permission clause sits only in the separate Falcon 180B license on the same page.

  • https://huggingface.co/api/models/tiiuae/Falcon3-10B-Base recorded 2026-09-29

    Hugging Face API record for "tiiuae/Falcon3-10B-Base" with gated and private both false and safetensors weight shards such as "model-00001-of-00005.safetensors"; the license tag is "license:other", so the Hub does not name the license and the tier is read from the TII text.

  • https://huggingface.co/blog/falcon3 recorded 2026-09-29

    Falcon 3 release post: "We conducted a single large-scale pretraining run on the 7B model, using 1024 H100 GPU chips, leveraging 14 trillion tokens featuring web, code, STEM, and curated high-quality and multilingual data" and "all models in the Falcon3 family are released under the" Falcon LLM license. The corpus is characterized but no dataset or training pipeline is linked.

  • https://huggingface.co/tiiuae/Falcon3-10B-Base/raw/main/README.md recorded 2026-09-29

    Falcon3-10B-Base model card front matter gives "license_name: falcon-llm-license" and "license_link: https://falconllm.tii.ae/falcon-terms-and-conditions.html"; the card states "License: TII Falcon-LLM License 2.0" and "Depth up-scaled from **Falcon3-7B-Base** with continual pretraining on 2 Teratokens". Its only code is a transformers inference snippet, and the Technical Report section reads "Coming soon....".

  • https://huggingface.co/tiiuae/Falcon3-7B-Base recorded 2026-09-29

    Falcon3-7B-Base model page, license tag falcon-llm-license, stating "Pretrained on 14 Teratokens of datasets comprising of web, code, STEM, high quality and mutlilingual data" and "License: TII Falcon-LLM License 2.0"; the embedded repo data records gated as false. Searched for agree to share and access request - none.

  • https://www.tii.ae/news/falcon-3-uaes-technology-innovation-institute-launches-worlds-most-powerful-small-ai-models recorded 2026-09-29

    TII launch announcement: "The Falcon 3 series include four model sizes" (1B, 3B, 7B and 10B) and "Falcon 3 is licensed under TII Falcon License, the permissive Apache 2.0-based software license that includes an acceptable use policy".

Adoption

1 high confidence
1.0

The single declared artifact, Falcon3-10B-Base, is scored on measured Hugging Face downloads rather than a reported-traction figure, since a direct count is the stronger signal when one exists. That count confirms what the Falcon 3 family's own description already says: relevance has faded against newer small language models.

Capability

2 high confidence
2.0

Falcon3-10B-Base scores 73.1 on MMLU and 42.5 on MMLU-Pro, and the 7B base 67.4 and 39.2. Those results were competitive among models under 13B when Falcon 3 was released, but sit well below the 2026 frontier.

  • https://huggingface.co/api/models?author=tiiuae&sort=createdAt&direction=-1&limit=30 recorded 2026-09-29

    Hugging Face listing of the 30 newest tiiuae repositories, newest first, headed by "tiiuae/Falcon3-10B-Base-1.58bit-prequantized" and including "tiiuae/Falcon-H1R-7B" and "tiiuae/Falcon-H1-Tiny-90M-Base"; no new Falcon 3 base checkpoint or larger Falcon base model appears since the recorded figures.

  • https://huggingface.co/blog/falcon3 recorded 2026-09-29

    Falcon 3 release post: "with scores of 67.4/39.2 (MMLU/MMLU-PRO) for Falcon3-7B-Base and 73.1/42.5 (MMLU/MMLU-PRO) for Falcon3-10B-Base respectively" and "Falcon3-10B-Base achieves 22.9 on MATH-Lvl5 and 83.0 on GSM8K".

  • https://huggingface.co/tiiuae/Falcon3-10B-Base/raw/main/README.md recorded 2026-09-29

    Falcon3-10B-Base model card benchmark table, internal lm-evaluation-harness raw scores: "We report in the following table our internal pipeline benchmarks." Its comparison set is Gemma2-9B, Yi1.5-9B and "Mistral-Nemo-Base-2407 (12B)", all 2024 releases.

Verified 2026-09-27