Welcome PaliGemma 2 β New vision language models by Google
β’
118
Merve Noyan
merve
AI & ML interests
VLMs, vision & co
Recent Activity
posted
an
update
1 day ago
QwQ can see π₯
Qwen team released QvQ, a large vision LM with reasoning π±
it outperforms proprietary VLMs on several benchmarks, comes with open weights and a demo!
Check them out β¬οΈ
Demo https://huggingface.co./spaces/Qwen/QVQ-72B-preview
Model https://huggingface.co./Qwen/QVQ-72B-Preview
Read more https://qwenlm.github.io/blog/qvq-72b-preview/
Congratulations @JustinLin610 and team!
updated
a model
5 days ago
merve/colpali_ufo
Articles
Organizations
merve's activity
posted
an
update
1 day ago
Post
1150
QwQ can see π₯
Qwen team released QvQ, a large vision LM with reasoning π±
it outperforms proprietary VLMs on several benchmarks, comes with open weights and a demo!
Check them out β¬οΈ
Demo Qwen/QVQ-72B-preview
Model Qwen/QVQ-72B-Preview
Read more https://qwenlm.github.io/blog/qvq-72b-preview/
Congratulations @JustinLin610 and team!
Qwen team released QvQ, a large vision LM with reasoning π±
it outperforms proprietary VLMs on several benchmarks, comes with open weights and a demo!
Check them out β¬οΈ
Demo Qwen/QVQ-72B-preview
Model Qwen/QVQ-72B-Preview
Read more https://qwenlm.github.io/blog/qvq-72b-preview/
Congratulations @JustinLin610 and team!
reacted to
AdinaY's
post with π₯
1 day ago
Post
1036
QvQ-72B-Previewπ an open weight model for visual reasoning just released by Alibaba_Qwen team
Qwen/qvq-676448c820912236342b9888
β¨ Combines visual understanding & language reasoning.
β¨ Scores 70.3 on MMMU
β¨ Outperforms Qwen2-VL-72B-Instruct in complex problem-solving
Qwen/qvq-676448c820912236342b9888
β¨ Combines visual understanding & language reasoning.
β¨ Scores 70.3 on MMMU
β¨ Outperforms Qwen2-VL-72B-Instruct in complex problem-solving
posted
an
update
7 days ago
Post
2330
Aya by Cohere For AI can now see! π
C4AI community has built Maya 8B, a new open-source multilingual VLM built on SigLIP and Aya 8B π± works on 8 languages! π£οΈ
The authors extend Llava dataset using Aya's translation capabilities with 558k examples!
ry it here kkr5155/maya_demo
Dataset maya-multimodal/pretrain
Model maya-multimodal/maya π
kudos @nahidalam and team
C4AI community has built Maya 8B, a new open-source multilingual VLM built on SigLIP and Aya 8B π± works on 8 languages! π£οΈ
The authors extend Llava dataset using Aya's translation capabilities with 558k examples!
ry it here kkr5155/maya_demo
Dataset maya-multimodal/pretrain
Model maya-multimodal/maya π
kudos @nahidalam and team
- 1 reply
posted
an
update
8 days ago
Post
2905
Apollo is a new family of open-source video language models by Meta, where 3B model outperforms most 7B models and 7B outperforms most 30B models π§Ά
β¨ the models come in 1.5B https://huggingface.co./Apollo-LMMs/Apollo-1_5B-t32, 3B https://huggingface.co./Apollo-LMMs/Apollo-3B-t32 and 7B https://huggingface.co./Apollo-LMMs/Apollo-7B-t32 with A2.0 license, based on Qwen1.5 & Qwen2
β¨ the authors also release a benchmark dataset https://huggingface.co./spaces/Apollo-LMMs/ApolloBench
The paper has a lot of experiments (they trained 84 models!) about what makes the video LMs work β―οΈ
Try the demo for best setup here https://huggingface.co./spaces/Apollo-LMMs/Apollo-3B
they evaluate sampling strategies, scaling laws for models and datasets, video representation and more!
> The authors find out that whatever design decision was applied to small models also scale properly when the model and dataset are scaled π scaling dataset has diminishing returns for smaller models
> They evaluate frame sampling strategies, and find that FPS sampling is better than uniform sampling, and they find 8-32 tokens per frame optimal
> They also compare image encoders, they try a variation of models from shape optimized SigLIP to DINOv2
they find google/siglip-so400m-patch14-384 to be most powerful π₯
> they also compare freezing different parts of models, training all stages with some frozen parts give the best yield
They eventually release three models, where Apollo-3B outperforms most 7B models and Apollo 7B outperforms 30B models π₯
β¨ the models come in 1.5B https://huggingface.co./Apollo-LMMs/Apollo-1_5B-t32, 3B https://huggingface.co./Apollo-LMMs/Apollo-3B-t32 and 7B https://huggingface.co./Apollo-LMMs/Apollo-7B-t32 with A2.0 license, based on Qwen1.5 & Qwen2
β¨ the authors also release a benchmark dataset https://huggingface.co./spaces/Apollo-LMMs/ApolloBench
The paper has a lot of experiments (they trained 84 models!) about what makes the video LMs work β―οΈ
Try the demo for best setup here https://huggingface.co./spaces/Apollo-LMMs/Apollo-3B
they evaluate sampling strategies, scaling laws for models and datasets, video representation and more!
> The authors find out that whatever design decision was applied to small models also scale properly when the model and dataset are scaled π scaling dataset has diminishing returns for smaller models
> They evaluate frame sampling strategies, and find that FPS sampling is better than uniform sampling, and they find 8-32 tokens per frame optimal
> They also compare image encoders, they try a variation of models from shape optimized SigLIP to DINOv2
they find google/siglip-so400m-patch14-384 to be most powerful π₯
> they also compare freezing different parts of models, training all stages with some frozen parts give the best yield
They eventually release three models, where Apollo-3B outperforms most 7B models and Apollo 7B outperforms 30B models π₯
- 3 replies
posted
an
update
13 days ago
Post
1669
A complete RAG pipeline includes a reranker, which ranks the documents to find the best document π
Same goes for multimodal RAG, multimodal rerankers which we can integrate to multimodal RAG pipelines!
Learn how to build a complete multimodal RAG pipeline with vidore/colqwen2-v1.0 as retriever, lightonai/MonoQwen2-VL-v0.1 as reranker, Qwen/Qwen2-VL-7B-Instruct as VLM in this notebook that runs on a GPU as small as L4 π₯ https://huggingface.co./learn/cookbook/multimodal_rag_using_document_retrieval_and_reranker_and_vlms
Same goes for multimodal RAG, multimodal rerankers which we can integrate to multimodal RAG pipelines!
Learn how to build a complete multimodal RAG pipeline with vidore/colqwen2-v1.0 as retriever, lightonai/MonoQwen2-VL-v0.1 as reranker, Qwen/Qwen2-VL-7B-Instruct as VLM in this notebook that runs on a GPU as small as L4 π₯ https://huggingface.co./learn/cookbook/multimodal_rag_using_document_retrieval_and_reranker_and_vlms
posted
an
update
17 days ago
Post
5505
This week in open-source AI was insane π€ A small recapπΊπ»
merve/dec-6-releases-67545caebe9fc4776faac0a3
Multimodal πΌοΈ
> Google shipped a PaliGemma 2, new iteration of PaliGemma with more sizes: 3B, 10B and 28B, with pre-trained and captioning variants π
> OpenGVLab released InternVL2, seven new vision LMs in different sizes, with sota checkpoint with MIT license β¨
> Qwen team at Alibaba released the base models of Qwen2VL models with 2B, 7B and 72B ckpts
LLMs π¬
> Meta released a new iteration of Llama 70B, Llama3.2-70B trained further
> EuroLLM-9B-Instruct is a new multilingual LLM for European languages with Apache 2.0 license π₯
> Dataset: CohereForAI released GlobalMMLU, multilingual version of MMLU with 42 languages with Apache 2.0 license
> Dataset: QwQ-LongCoT-130K is a new dataset to train reasoning models
> Dataset: FineWeb2 just landed with multilinguality update! π₯ nearly 8TB pretraining data in many languages!
Image/Video Generation πΌοΈ
> Tencent released HunyuanVideo, a new photorealistic video generation model
> OminiControl is a new editing/control framework for image generation models like Flux
Audio π
> Indic-Parler-TTS is a new text2speech model made by community
Multimodal πΌοΈ
> Google shipped a PaliGemma 2, new iteration of PaliGemma with more sizes: 3B, 10B and 28B, with pre-trained and captioning variants π
> OpenGVLab released InternVL2, seven new vision LMs in different sizes, with sota checkpoint with MIT license β¨
> Qwen team at Alibaba released the base models of Qwen2VL models with 2B, 7B and 72B ckpts
LLMs π¬
> Meta released a new iteration of Llama 70B, Llama3.2-70B trained further
> EuroLLM-9B-Instruct is a new multilingual LLM for European languages with Apache 2.0 license π₯
> Dataset: CohereForAI released GlobalMMLU, multilingual version of MMLU with 42 languages with Apache 2.0 license
> Dataset: QwQ-LongCoT-130K is a new dataset to train reasoning models
> Dataset: FineWeb2 just landed with multilinguality update! π₯ nearly 8TB pretraining data in many languages!
Image/Video Generation πΌοΈ
> Tencent released HunyuanVideo, a new photorealistic video generation model
> OminiControl is a new editing/control framework for image generation models like Flux
Audio π
> Indic-Parler-TTS is a new text2speech model made by community
posted
an
update
18 days ago
Post
1500
New InternVL drop with a state-of-the-art 78B vision language model with MIT license π₯ https://huggingface.co./collections/OpenGVLab/internvl-25-673e1019b66e2218f68d7c1c
The release comes with seven new vision LMs based on InternViT 300M/6B and Qwen2.5 (0.5B, 3B, 32B, 72B) and InternLM2 (8B, 7B, 20B) in different sizes
78B model is of InternViT 6B and Qwen2.5-72B Instruct, can accomplish variety of tasks π Try here OpenGVLab/InternVL
The release comes with seven new vision LMs based on InternViT 300M/6B and Qwen2.5 (0.5B, 3B, 32B, 72B) and InternLM2 (8B, 7B, 20B) in different sizes
78B model is of InternViT 6B and Qwen2.5-72B Instruct, can accomplish variety of tasks π Try here OpenGVLab/InternVL
posted
an
update
23 days ago
Post
2628
small but mighty π₯
you can fine-tune SmolVLM on an L4 with batch size of 4 and it will only take 16.4 GB VRAM π«°π» also with gradient accumulation simulated batch size is 16 β¨
I made a notebook that includes all the goodies: QLoRA, gradient accumulation, gradient checkpointing with explanations on how they work π https://github.com/huggingface/smollm/blob/main/finetuning/Smol_VLM_FT.ipynb
you can fine-tune SmolVLM on an L4 with batch size of 4 and it will only take 16.4 GB VRAM π«°π» also with gradient accumulation simulated batch size is 16 β¨
I made a notebook that includes all the goodies: QLoRA, gradient accumulation, gradient checkpointing with explanations on how they work π https://github.com/huggingface/smollm/blob/main/finetuning/Smol_VLM_FT.ipynb
posted
an
update
23 days ago
Post
2864
Last week we were blessed with open-source models! A recap π
merve/nov-29-releases-674ccc255a57baf97b1e2d31
πΌοΈ Multimodal
> At Hugging Face we released SmolVLM, a performant and efficient smol vision language model π
> Show Lab released ShowUI-2B: new vision-language-action model to build GUI/web automation agents π€
> Rhymes AI has released the base model of Aria: Aria-Base-64K and Aria-Base-8K with their respective context length
> ViDoRe team released ColSmolVLM: A new ColPali-like retrieval model based on SmolVLM
> Dataset: Llava-CoT-o1-Instruct: new dataset labelled using Llava-CoT multimodal reasoning modelπ
> Dataset: LLaVA-CoT-100k dataset used to train Llava-CoT released by creators of Llava-CoT π
π¬ LLMs
> Qwen team released QwQ-32B-Preview, state-of-the-art open-source reasoning model, broke the internet π₯
> AliBaba has released Marco-o1, a new open-source reasoning model π₯
> NVIDIA released Hymba 1.5B Base and Instruct, the new state-of-the-art SLMs with hybrid architecture (Mamba + transformer)
β―οΈ Image/Video Generation
> Qwen2VL-Flux: new image generation model based on Qwen2VL image encoder, T5 and Flux for generation
> Lightricks released LTX-Video, a new DiT-based video generation model that can generate 24 FPS videos at 768x512 res β―οΈ
> Dataset: Image Preferences is a new image generation preference dataset made with DIBT community effort of Argilla π·οΈ
Audio
> OuteAI released OuteTTS-0.2-500M new multilingual text-to-speech model based on Qwen-2.5-0.5B trained on 5B audio prompt tokens
merve/nov-29-releases-674ccc255a57baf97b1e2d31
πΌοΈ Multimodal
> At Hugging Face we released SmolVLM, a performant and efficient smol vision language model π
> Show Lab released ShowUI-2B: new vision-language-action model to build GUI/web automation agents π€
> Rhymes AI has released the base model of Aria: Aria-Base-64K and Aria-Base-8K with their respective context length
> ViDoRe team released ColSmolVLM: A new ColPali-like retrieval model based on SmolVLM
> Dataset: Llava-CoT-o1-Instruct: new dataset labelled using Llava-CoT multimodal reasoning modelπ
> Dataset: LLaVA-CoT-100k dataset used to train Llava-CoT released by creators of Llava-CoT π
π¬ LLMs
> Qwen team released QwQ-32B-Preview, state-of-the-art open-source reasoning model, broke the internet π₯
> AliBaba has released Marco-o1, a new open-source reasoning model π₯
> NVIDIA released Hymba 1.5B Base and Instruct, the new state-of-the-art SLMs with hybrid architecture (Mamba + transformer)
β―οΈ Image/Video Generation
> Qwen2VL-Flux: new image generation model based on Qwen2VL image encoder, T5 and Flux for generation
> Lightricks released LTX-Video, a new DiT-based video generation model that can generate 24 FPS videos at 768x512 res β―οΈ
> Dataset: Image Preferences is a new image generation preference dataset made with DIBT community effort of Argilla π·οΈ
Audio
> OuteAI released OuteTTS-0.2-500M new multilingual text-to-speech model based on Qwen-2.5-0.5B trained on 5B audio prompt tokens
posted
an
update
28 days ago
Post
2161
The authors of ColPali trained a retrieval model based on SmolVLM π€
vidore/colsmolvlm-alpha
TLDR;
- ColSmolVLM performs better than ColPali and DSE-Qwen2 on all English tasks
- ColSmolVLM is more memory efficient than ColQwen2 π
TLDR;
- ColSmolVLM performs better than ColPali and DSE-Qwen2 on all English tasks
- ColSmolVLM is more memory efficient than ColQwen2 π
posted
an
update
29 days ago
Post
3870
Small yet mighty! π«
We are releasing SmolVLM: a new 2B small vision language made for on-device use, fine-tunable on consumer GPU, immensely memory efficient π€
We release three checkpoints under Apache 2.0: SmolVLM-Instruct, SmolVLM-Synthetic and SmolVLM-Base HuggingFaceTB/smolvlm-6740bd584b2dcbf51ecb1f39
Learn more from our blog here: huggingface.co/blog/smolvlm
This release comes with a demo, fine-tuning code, MLX integration and TRL integration for DPO π
Try the demo: HuggingFaceTB/SmolVLM
Fine-tuning Recipe: https://github.com/huggingface/smollm/blob/main/finetuning/Smol_VLM_FT.ipynb
Also TRL integration for DPO π
We are releasing SmolVLM: a new 2B small vision language made for on-device use, fine-tunable on consumer GPU, immensely memory efficient π€
We release three checkpoints under Apache 2.0: SmolVLM-Instruct, SmolVLM-Synthetic and SmolVLM-Base HuggingFaceTB/smolvlm-6740bd584b2dcbf51ecb1f39
Learn more from our blog here: huggingface.co/blog/smolvlm
This release comes with a demo, fine-tuning code, MLX integration and TRL integration for DPO π
Try the demo: HuggingFaceTB/SmolVLM
Fine-tuning Recipe: https://github.com/huggingface/smollm/blob/main/finetuning/Smol_VLM_FT.ipynb
Also TRL integration for DPO π
posted
an
update
about 1 month ago
Post
2579
What a week! A recap for everything you missed βοΈ
merve/nov-22-releases-673fbbcfc1c97c4f411def07
Multimodal β¨
> Mistral AI
released Pixtral 124B, a gigantic open vision language model
> Llava-CoT (formerly known as Llava-o1) was released, a multimodal reproduction of o1 model by PKU
> OpenGVLab released MMPR: a new multimodal reasoning dataset
> Jina has released Jina-CLIP-v2 0.98B multilingual multimodal embeddings
> Apple released new SotA vision encoders AIMv2
LLMs π¦
> AllenAI dropped a huge release of models, datasets and scripts for TΓΌlu, a family of models based on Llama 3.1 aligned with SFT, DPO and a new technique they have developed called RLVR
> Jina has released embeddings-v3: new multilingual embeddings with longer context
> Hugging Face released SmolTalk: synthetic dataset used to align SmolLM2 using supervised fine-tuning
> Microsoft released orca-agentinstruct-1M-v1: a gigantic instruction dataset of 1M synthetic instruction pairs
Image Generation πΌοΈ
> Black Forest Labs released Flux 1. tools: four new models for different image modifications and two LoRAs to do image conditioning and better steer generations
Lastly Hugging Face released a new library Observers: a lightweight SDK for monitoring interactions with AI APIs and easily store and browse them π
$ pip install observers
merve/nov-22-releases-673fbbcfc1c97c4f411def07
Multimodal β¨
> Mistral AI
released Pixtral 124B, a gigantic open vision language model
> Llava-CoT (formerly known as Llava-o1) was released, a multimodal reproduction of o1 model by PKU
> OpenGVLab released MMPR: a new multimodal reasoning dataset
> Jina has released Jina-CLIP-v2 0.98B multilingual multimodal embeddings
> Apple released new SotA vision encoders AIMv2
LLMs π¦
> AllenAI dropped a huge release of models, datasets and scripts for TΓΌlu, a family of models based on Llama 3.1 aligned with SFT, DPO and a new technique they have developed called RLVR
> Jina has released embeddings-v3: new multilingual embeddings with longer context
> Hugging Face released SmolTalk: synthetic dataset used to align SmolLM2 using supervised fine-tuning
> Microsoft released orca-agentinstruct-1M-v1: a gigantic instruction dataset of 1M synthetic instruction pairs
Image Generation πΌοΈ
> Black Forest Labs released Flux 1. tools: four new models for different image modifications and two LoRAs to do image conditioning and better steer generations
Lastly Hugging Face released a new library Observers: a lightweight SDK for monitoring interactions with AI APIs and easily store and browse them π
$ pip install observers
- 3 replies
posted
an
update
about 1 month ago
Post
1495
Apple released AIMv2 π a family of state-of-the-art open-set vision encoders
apple/aimv2-6720fe1558d94c7805f7688c
> like CLIP, but add a decoder and train on autoregression π€―
> 19 open models come in 300M, 600M, 1.2B, 2.7B with resolutions of 224, 336, 448
> Load and use with π€ transformers
apple/aimv2-6720fe1558d94c7805f7688c
> like CLIP, but add a decoder and train on autoregression π€―
> 19 open models come in 300M, 600M, 1.2B, 2.7B with resolutions of 224, 336, 448
> Load and use with π€ transformers
posted
an
update
about 1 month ago
Post
3103
your hugging face profile now has your recent activities π€
posted
an
update
about 1 month ago
Post
1810
π€ transformers pipelines now support vision language models for easy local inference π«°π»
h/t @yonigozlan for shipping this π©π
you can also use inference API to infer hosted vision LMs (via Python, JS and cURL) https://huggingface.co./docs/api-inference/en/tasks/image-text-to-text
h/t @yonigozlan for shipping this π©π
you can also use inference API to infer hosted vision LMs (via Python, JS and cURL) https://huggingface.co./docs/api-inference/en/tasks/image-text-to-text
reacted to
sayakpaul's
post with β€οΈ
about 1 month ago
Post
2601
It's been a while we shipped native quantization support in
We currently support
This post is just a reminder of what's possible:
1. Loading a model with a quantization config
2. Saving a model with quantization config
3. Loading a pre-quantized model
4.
5. Training and loading LoRAs into quantized checkpoints
Docs:
https://huggingface.co./docs/diffusers/main/en/quantization/bitsandbytes
diffusers π§¨We currently support
bistandbytes as the official backend but using others like torchao is already very simple. This post is just a reminder of what's possible:
1. Loading a model with a quantization config
2. Saving a model with quantization config
3. Loading a pre-quantized model
4.
enable_model_cpu_offload()5. Training and loading LoRAs into quantized checkpoints
Docs:
https://huggingface.co./docs/diffusers/main/en/quantization/bitsandbytes
- 1 reply
reacted to
BlinkDL's
post with π₯
about 1 month ago
Post
3961
RWKV-6-world-v3 (+3.1T tokens) is our best multilingual 7B model as of now:
BlinkDL/rwkv-6-world
It's 100% RNN and attention-free. MMLU 54.2% (previous world-v2.1 = 47.9%. note: without eval-boosting tricks such as annealing).
RWKV-7-world-v4 soon :)
It's 100% RNN and attention-free. MMLU 54.2% (previous world-v2.1 = 47.9%. note: without eval-boosting tricks such as annealing).
RWKV-7-world-v4 soon :)
reacted to
davidberenstein1957's
post with π₯π
about 1 month ago
Post
1960
For anyone who struggles with NER or information extraction with LLM.
We showed an efficient workflow for token classification including zero-shot suggestions and model fine-tuning with Argilla, GliNER, the NuMind NuExtract LLM and SpanMarker. @argilla
Video: https://youtu.be/JvLpaYgNd84?feature=shared
Notebooks and slides included to try it yourself π
We showed an efficient workflow for token classification including zero-shot suggestions and model fine-tuning with Argilla, GliNER, the NuMind NuExtract LLM and SpanMarker. @argilla
Video: https://youtu.be/JvLpaYgNd84?feature=shared
Notebooks and slides included to try it yourself π
reacted to
erikkaum's
post with π
about 1 month ago
Post
1704
A while ago I started experimenting with compiling the Python interpreter to WASM.
To build a secure, fast, and lightweight sandbox for code execution β ideal for running LLM-generated Python code.
- Send code simply as a POST request
- 1-2ms startup times
Hack away:
https://github.com/ErikKaum/runner
To build a secure, fast, and lightweight sandbox for code execution β ideal for running LLM-generated Python code.
- Send code simply as a POST request
- 1-2ms startup times
Hack away:
https://github.com/ErikKaum/runner