Hugging Face

SmolLM3

Hugging Face

SmolLM3 is a Hugging Face open-weight model family. The 3B successor adds optional reasoning, six natively supported languages, and long-context capabilities. It is trained at 64K context and supports extension to 128K through YaRN. The public release includes training recipes, data mixture, and base/instruction weights for reproducible small-model research. Apache-2.0 licensing supports local deployment and adaptation. Model outputs can hallucinate and reflect training biases; multimodal tiers can also misread fine visual details.

SmolVLM2

Hugging Face

SmolVLM2 is a Hugging Face open-weight model family. The second vision-language generation adds video understanding to compact image-and-text models. It combines SigLIP visual features with SmolLM2 text decoding and trains on video and image-language datasets. Several size tiers target video-capable local devices; frame sampling and visual tokens determine effective media capacity. Apache-2.0 licensing supports local deployment and adaptation. Model outputs can hallucinate and reflect training biases; multimodal tiers can also misread fine visual details.

SmolVLM

Hugging Face

SmolVLM is a Hugging Face open-weight model family. This small vision-language family pairs a compact text decoder with a SigLIP vision encoder using an Idefics-style connector. It supports image questions, document interpretation, and multi-image chat with modest memory needs. The later 256M and 500M tiers are variants of the original named release. Apache-2.0 licensing supports local deployment and adaptation. Model outputs can hallucinate and reflect training biases; multimodal tiers can also misread fine visual details.

SmolLM2

Hugging Face

SmolLM2 is a Hugging Face open-weight model family. The second generation extends training data and token budgets for small-device assistants. It improves instruction following, mathematics, code, and concise chat, with base and instruction variants across three sizes. A later 16K instruction checkpoint is a long-context variant rather than an independently named generation. Apache-2.0 licensing supports local deployment and adaptation. Model outputs can hallucinate and reflect training biases; multimodal tiers can also misread fine visual details.

Idefics3

Hugging Face

Idefics3 is a Hugging Face open-weight multimodal language-model release. Idefics3 is the third open vision-language generation, documented with a reproducible model-building study and open multimodal datasets. It combines SigLIP visual features with a Llama 3.1 8B instruction backbone and supports image questions, interleaved multi-image conversations, document analysis, and textual output. The card declares Apache 2.0 for its release while identifying the Llama backbone; users should consider the terms of that component. The date records the public technical report rather than assuming an earlier repository timestamp was a release. Perceptual interpretation and extracted text can be wrong, and performance depends on image resolution and task framing.

SmolLM

Hugging Face

SmolLM is a Hugging Face open-weight model family. The first small-model family uses carefully curated educational, synthetic, and web data to support inexpensive local language generation. Base and instruction sizes run on relatively modest hardware but are not substitutes for larger models on demanding reasoning. Apache-2.0 licensing supports local deployment and adaptation. Model outputs can hallucinate and reflect training biases; multimodal tiers can also misread fine visual details.

Idefics2

Hugging Face

Idefics2 is a Hugging Face open-weight multimodal language-model release. Idefics2 improves visual reasoning and document understanding in a smaller eight-billion-parameter model. It combines a SigLIP visual encoder, a Mistral language decoder, and a connector that efficiently handles images at their native aspect ratios. Interleaved images and text support visual question answering, image comparison, and document extraction. Public base, instruction, and conversational checkpoints share this named generation. Perceptual interpretation and extracted text can be wrong, and performance depends on image resolution and task framing.

IDEFICS

Hugging Face

IDEFICS is a Hugging Face open-weight multimodal language-model release. The initial IDEFICS family openly reproduces Flamingo-style interleaved image/text modeling at 9B and 80B scale. A CLIP vision encoder connects through newly trained components to a Llama text backbone. Base and instruction variants support multi-image questions, visual descriptions, and conversational text generation. Additional weights are MIT-licensed but the original Llama backbone retains non-commercial research terms. Perceptual interpretation and extracted text can be wrong, and performance depends on image resolution and task framing.

BLOOMZ

Hugging Face

BLOOMZ is BigScience’s crosslingual multitask adaptation of the BLOOM language-model family. Training on xP3 combines multilingual instructions and tasks to improve zero-shot generalization across both task types and languages, making it distinct from BLOOM’s base continuation behavior. Released sizes span small local checkpoints through the 176B tier, with different prompting and multilingual training variants. The accompanying mT0 uses a separate encoder-decoder backbone and is not treated as a BLOOM checkpoint. The public paper releases models and training code; results depend on task instructions and language coverage.

BLOOM

Hugging Face

BLOOM is the BigScience collaboration’s openly released 176B multilingual language model, developed with Hugging Face and an international research community. It is a dense autoregressive Transformer trained on the ROOTS multilingual corpus and released with research documentation and training artifacts. The model supports forty-six natural languages and thirteen programming languages, with smaller size checkpoints belonging to the same family. It is primarily a continuation model rather than a modern instruction-tuned assistant. The custom Responsible AI License includes use restrictions, so open weights should not be confused with unrestricted Apache licensing.