AI Radar

Hugging Face models

Hugging Face is a company founded in 2016, incorporated in the US as Hugging Face, Inc., with its main EU establishment in Paris. It runs a platform for hosting and sharing machine learning models and datasets, offers courses and tools, and says it works to democratise AI and machine learning. It also releases open models, including the SmolLM and SmolVLM families.

Chronology

  1. 2025

    3 models
    1. Released
      SmolLM · Language, Reasoning · Milestone SmolLM3

      SmolLM3 is a 3B-parameter open-weight language model from Hugging Face, released under Apache 2.0 as a base model and an instruct model with switchable extended thinking. It natively supports six…

      • The instruct model adds an extended-thinking mode, on by default and switched with /think and /no_think flags.
      • Natively supports six languages: English, French, Spanish, German, Italian and Portuguese.
      Primary source Available · no news linked
    2. Released
      Smol Models · Language, Multimodal, Robotics · Milestone SmolVLA

      SmolVLA is a 450M-parameter open-weight vision-language-action model for robotics from Hugging Face's LeRobot project. It pairs SmolVLM2 with an action expert trained with flow matching to turn…

      • Adds robot state input and continuous action output, generated by a flow-matching action expert on top of the SmolVLM2 backbone.
      Primary source Available · no news linked
    3. Released
      SmolVLM · Language, Multimodal · Milestone SmolVLM2

      SmolVLM2 extends Hugging Face's SmolVLM vision-language models to video, taking video, images and text as input and producing text. It was released with open weights under Apache 2.0 in 2.2B, 500M…

      • Accepts video input in addition to images and text.
      • Training data adds video datasets such as FineVideo to The Cauldron and Docmatix.
      Primary source Available · no news linked

  2. 2024

    7 models
    1. Released
      SmolVLM · Language, Multimodal · Milestone SmolVLM

      SmolVLM is a 2B open-weight vision-language model from Hugging Face that takes sequences of images and text and produces text, released as Base, Synthetic and Instruct checkpoints under Apache 2.0…

      • Uses SmolLM2 1.7B instead of Llama 3.1 8B as the language backbone, within the Idefics3 architecture.
      • Compresses visual information 9x with pixel shuffle, compared with 4x in Idefics3.
      Primary source Available · no news linked
    2. Day not given
      SmolLM · Language · Milestone SmolLM2

      SmolLM2 is the second generation of Hugging Face's compact open-weight language models, released in 135M, 360M and 1.7B sizes under Apache 2.0. Hugging Face describes them as light enough to run…

      • The 1.7B model was trained on 11 trillion tokens, up from 1 trillion for SmolLM-1.7B.
      • Context length of the 1.7B model extended from 2k to 8k tokens late in pre-training.
      Primary source Available · no news linked
    3. Day not given
      Idefics · Language, Multimodal · Milestone Idefics3

      Idefics3 is an 8B open-weight multimodal model from Hugging Face that takes interleaved images and text and produces text, built on Meta's Llama 3.1 8B Instruct and Google's SigLIP-SO400M and…

      • Uses Llama 3.1 Instruct instead of Mistral-7B as the language model.
      • Encodes each image tile of up to 364x364 pixels with 169 visual tokens; larger images are split into tiles that are encoded separately.
      Primary source Available · no news linked
    4. Released
      SmolLM · Language · Milestone SmolLM

      SmolLM is a family of small open-weight language models from Hugging Face in 135M, 360M and 1.7B sizes, each released as a base and an instruction-tuned checkpoint under Apache 2.0. Hugging Face…

      Primary source Available · no news linked
    5. Released
      Idefics · Language, Multimodal · Milestone Idefics2

      Idefics2 is an 8B multimodal model from Hugging Face that takes interleaved text and images and generates text, published as open weights under Apache 2.0 in base, instruction-tuned and chatty…

      • A single 8B size instead of 9B and 80B; Hugging Face calls it 10x smaller than the 80B model.
      • Replaces gated cross-attention with Perceiver pooling and an MLP projection, and uses Mistral-7B-v0.1 and SigLIP so400m as backbones.
      Primary source Available · no news linked
    6. Released
      Zephyr · Language Zephyr 141B-A35B

      Zephyr 141B-A35B is a Zephyr chat model from Hugging Face, fine-tuned from the Mixture of Experts model Mixtral 8x22B v0.1 and published as open weights under Apache 2.0. The card says Argilla…

      • A Mixture of Experts model with 141B total and 39B active parameters, against 7B for Zephyr 7B Gemma.
      • Built on the Mixture of Experts model Mixtral 8x22B v0.1 as base model instead of Gemma 7B.
      Primary source Available · no news linked
    7. Released
      Zephyr · Language Zephyr 7B Gemma

      Zephyr 7B Gemma is the third model in Hugging Face's Zephyr series, a 7B chat assistant fine-tuned from Google's Gemma 7B and published as open weights under the Gemma Terms of Use. It was…

      • Built on Google's Gemma 7B as base model instead of Mistral 7B v0.1.
      • Fine-tuned on the DEITA 10K dataset and aligned with DPO on argilla/dpo-mix-7k, about 7k prompts with completions scored by GPT-4.
      Primary source Available · no news linked

  3. 2023

    3 models
    1. Day not given
      Zephyr · Language · Milestone Zephyr 7B Alpha

      Zephyr 7B Alpha is the first model in Hugging Face's Zephyr series of chat assistants, a 7B fine-tune of Mistral 7B v0.1 published as open weights under the MIT licence. Hugging Face says it removed…

      Primary source Available · no news linked
    2. Day not given
      Zephyr · Language Zephyr 7B β

      Zephyr 7B β is the second model in Hugging Face's Zephyr series, a 7B chat assistant fine-tuned from Mistral 7B v0.1 and published as open weights under the MIT licence. Its technical report…

      • Trained on a filtered and preprocessed version of UltraChat and a binarized version of UltraFeedback, whose completions were scored by GPT-4.
      Primary source Available · no news linked
    3. Released
      Idefics · Language, Multimodal · Milestone IDEFICS

      IDEFICS is an open-weights visual language model from Hugging Face that takes interleaved images and text and produces text, released in 9B and 80B sizes, each as a base and an instruction-tuned…

      Primary source Available · no news linked

Lineage

  1. IDEFICS
    Successors
    Idefics2
    Based on
    LLaMA (Meta) · derived: other
  2. Zephyr 7B Alpha
    Successors
    Zephyr 7B β
    Based on
    Mistral 7B (Mistral AI) · derived: fine tune
  3. Zephyr 7B β
    Predecessors
    Zephyr 7B Alpha
    Successors
    Zephyr 7B Gemma
    Based on
    Mistral 7B (Mistral AI) · derived: fine tune
  4. Zephyr 7B Gemma
    Predecessors
    Zephyr 7B β
    Based on
    Gemma (Google) · derived: fine tune
  5. Zephyr 141B-A35B
    Predecessors
    Zephyr 7B Gemma
    Based on
    Mixtral 8x22B (Mistral AI) · derived: fine tune
  6. Idefics2
    Predecessors
    IDEFICS
    Successors
    Idefics3
    Based on
    Mistral 7B (Mistral AI) · derived: other
  7. SmolLM
    Successors
    SmolLM2
  8. Idefics3
    Predecessors
    Idefics2
    Based on
    Llama 3.1 (Meta) · derived: other
  9. SmolLM2
    Predecessors
    SmolLM
    Successors
    SmolLM3
    Variants and derived
    SmolVLM · derived: other
  10. SmolVLM
    Successors
    SmolVLM2
    Based on
    SmolLM2 · derived: other
  11. SmolVLM2
    Predecessors
    SmolVLM
    Variants and derived
    SmolVLA · derived: other
  12. SmolVLA
    Based on
    SmolVLM2 · derived: other
  13. SmolLM3
    Predecessors
    SmolLM2

Families

About the organisation

  • Description Hugging Face is a company founded in 2016, incorporated in the US as Hugging Face, Inc., with its main EU establishment in Paris. It runs a platform for hosting and sharing machine learning models and datasets, offers courses and tools, and says it works to democratise AI and machine learning. It also releases open models, including the SmolLM and SmolVLM families. Primary source[1] [2] [3] [4]

Company · https://huggingface.co · 13 models listed here

Sources

  1. [1] Hugging Face Response to Request for Information on the Development of an Artificial Intelligence Action Plan

    Other · Hugging Face · published

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://huggingface.co/datasets/huggingface/policy-docs/resolve/main/2025_Hugging_Face_Response_to_AI_Action_Plan.pdf PDF in the huggingface/policy-docs dataset on the Hugging Face Hub, linked from the blog post https://huggingface.co/blog/ai-action-wh-2025. The cover gives March 2025. The resolve URL answers 302 to a signed, expiring CDN address, so the stable Hub URL is kept. Facts used: the About Hugging Face section (community-driven U.S. company; hosts models and datasets; educational resources, courses and tooling) and the sentence saying it was founded in 2016.

  2. [2] Terms of Service

    Other · Hugging Face

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://huggingface.co/terms-of-service Living page; states Effective Date: September 15, 2022 and names the contracting party as Hugging Face, Inc., a Delaware corporation. HTML title: Terms of Service – Hugging Face.

  3. [3] Hugging Face Privacy Policy

    Other · Hugging Face

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://huggingface.co/privacy Living page; states Effective Date: March 28, 2023. Names Hugging Face SAS, headquartered at 9 rue des Colonnes, Paris, as the main establishment in the European Union. HTML title: Privacy Policy – Hugging Face.

  4. [4] huggingface/smollm

    Repository · Hugging Face

    Provenance
    Primary
    Availability
    Active
    Last checked
    Original ↗ https://github.com/huggingface/smollm Hugging Face GitHub repository; the README is headed Smol Models and describes a family of efficient, fully open models from Hugging Face for text and vision that can run on-device (SmolLM3, SmolLM2, SmolLM, SmolVLM). HTML title: GitHub - huggingface/smollm: Everything about the SmolLM and SmolVLM family of models. Living page.