We just shipped a blog on everything latest on vision language models, including 🤖 GUI agents, agentic VLMs, omni models 📑 multimodal RAG ⏯️ video LMs 🤏🏻 smol models ..and more! https://huggingface.co/blog/vlms-2025
Dropping some image classification models for content moderation, balancers, and classifiers trained on synthetic datasets—along with others based on datasets available on the Hub. Also loaded a few low-rank datasets for realistic gender portrait classification and document-type classifiers, all fine-tuned on the SigLIP-2 Patch-16 224 backbone. Models and datasets are listed below:
Well, here’s the updated version with the 20,000+ entry sampled dataset for Watermark Filter Content Moderation models incl. [Food25, Weather, Watermark, Marathi/Hindi Sign Language Detection], post-trained from the base models: sigLip2 patch16 224 — now with mixed aspect ratios for better performance and reduced misclassification. 🔥
💬 Qwen made it rain! They released Qwen3: new dense and MoE models ranging from 0.6B to 235B 🤯 as well as Qwen2.5-Omni, any-to-any model in 3B and 7B! > Microsoft AI released Phi4 reasoning models (that also come in mini and plus sizes) > NVIDIA released new CoT reasoning datasets 🖼️ > ByteDance released UI-TARS-1.5, native multimodal UI parsing agentic model > Meta released EdgeTAM, an on-device object tracking model (SAM2 variant) 🗣️ NVIDIA released parakeet-tdt-0.6b-v2, a smol 600M automatic speech recognition model > Nari released Dia, a 1.6B text-to-speech model > Moonshot AI released Kimi Audio, a new audio understanding, generation, conversation model 👩🏻💻 JetBrains released Melium models in base and SFT for coding > Tesslate released UIGEN-T2-7B, a new text-to-frontend-code model 🤩
When I heard the Reasoning Dataset Competition deadline was extended to 9 May, I knew I had time to get in one more entry. 🔥🔥🔥
With the rise of Vibe Coding, and the potential risks that are introduced by humans letting LLMs build their apps for them, lots of people are (rightfully) concerned about the safety of the code that is hitting prod.
In response to that, I'm happy to present my final submission to the Reasoning Dataset Competition and attempt to start benchmarking the ability of LLMs to identify unsafe and / or exploitable code by way of the CoSa (Code Safety) benchmark: ZennyKenny/cosa-benchmark-dataset
Currently a curated set of 200 examples, calibrated on OpenAI's standard issue models (GPT-4.1, o4 mini, and GPT-3.5 Turbo) as "baseline performance" (70% decile). Check it out and drop a ❤️ if you think it could be useful or hit the Community section with suggestions / critiques.
The new versions of Midjourney Mix adapters have been dropped in stranger zone hf. These adapters excel in studio lighting portraits and painterly styles, trained using the style of strangerzonehf/Flux-Midjourney-Mix2-LoRA. They leverage 24-bit colored synthetic images generated form midjourney v6 to achieve high-quality image reproducibility and support adaptable aspect ratios, using Flux.1 as the base model. 🥳
The best dimensions and inference settings for optimal results are as follows: A resolution of 1280 x 832 with a 3:2 aspect ratio is recommended for the best quality, while 1024 x 1024 with a 1:1 aspect ratio serves as the default option. For inference, the recommended number of steps ranges between 30 and 35 to achieve optimal output.
you can easily fine-tune, quantize, play with sota vision LM InternVL3 now 🔥 we have recently merged InternVL3 to Hugging Face transformers and released converted checkpoints 🤗
PSA for anyone using Nymbo/Nymbo_Theme or Nymbo/Nymbo_Theme_5 in a Gradio space ~
Both of these themes have been updated to fix some of the long-standing inconsistencies ever since the transition to Gradio v5. Textboxes are no longer bright green and in-line code is readable now! Both themes are now visually identical across versions.
If your space is already using one of these themes, you just need to restart your space to get the latest version. No code changes needed.
Dropping downstream tasks using newly initialized parameters and weights supports domain-specific image classification post-training, based on the SigLIP-2 models: Patch-16/224, Patch-16/256, and Patch-32/256. For more details, please refer to the respective model cards : 🤗
Meta released Llama Guard 4 and new Prompt Guard 2 models 🔥
Llama Guard 4 is a new model to filter model inputs/outputs both text-only and image 🛡️ use it before and after LLMs/VLMs! meta-llama/Llama-Guard-4-12B
The same way the advent of Adobe Illustrator has led to innovation in the way that creative professionals work, I earnestly believe that AI will do the same (contrary to the popular opinion that it represents some regression in the world of creatives).
@natalika and I were speaking about this topic and like most illustrators she has some understandable concerns about the spread of AI in her field. She also told me how much time she spends generating concept art that will never see the light of day in >98% of cases. 💡
To me, that sounded like a perfect opportunity to leverage image diffusion in a way that helps artists spend more time creating cool stuff rather than just malevolently mining their work and using it without credit. Using the Black Forest Labs base model FLUX, Replicate, and about $5 of H100 compute, I post-trained a LoRA adapter on a set of her images associated with one project she's working on and spun up an app with Hugging Face Spaces (and Zero GPU for the win).
Now, generating concept art in her particular style takes seconds instead of hours and when it's time to put the work into production, a human designer is still invaluable. And building it in the open hopefully inspires other use cases amongst other designers. 🖖
I've created a new dataset using the Algorithm of Thoughts architecture proposed by Sel et al. (2023) in a reasoning context. (paper: https://arxiv.org/pdf/2308.10379)
The dataset simulates the discovery phase of a fictitious VC firm called Reasoned Capital and, once expanded, can be used to create models which are able to make complex, subjective financial decisions based on different criteria.
The generation process encourages recursive problem-solving in increasingly complex prompts to encourage models to assess and reevaluate the conclusions and generated opinions of upstream models. Pretty neat stuff, and I'm not aware of this architecture being used in a reasoning context anywhere else.
Bringing out style-intermixing adapters for Flux.Dev, including Aura Glow, Fallen Ink Art, Cardboard Paper Arts, Black & White Expressions, and Glitter Gem Touch. For more details, visit the model card of the LoRA. 🥳
The best dimensions and inference settings for optimal results are as follows: A resolution of 1280 x 832 with a 3:2 aspect ratio is recommended for the best quality, while 1024 x 1024 with a 1:1 aspect ratio serves as the default option. For inference, the recommended number of steps ranges between 30 and 35 to achieve optimal output.
I'd be interested to hear the community's thoughts on the applications of AI in the military. Especially in the wargaming space.
This is something that feels inevitable (and realistically, probably already in progress). Doesn't it make sense for us to have an understanding of the mechanics of such processes? Surely they will never be open source.
Meta dropped swiss army knives for vision with A2.0 license 👏 > image/video encoders for vision language modelling and spatial understanding (object detection etc) 👏 > The vision LM outperforms InternVL3 and Qwen2.5VL 👏 > They also release gigantic video and image datasets
The authors attempt to come up with single versatile vision encoder to align on diverse set of tasks.
They trained Perception Encoder (PE) Core: a new state-of-the-art family of vision encoders that can be aligned for both vision-language and spatial tasks. For zero-shot image tasks, it outperforms latest sota SigLIP2 👏
> Among fine-tuned ones, first one is PE-Spatial. It's a model to detect bounding boxes, segmentation, depth estimation and it outperforms all other models 😮
> Second one is PLM, Perception Language Model, where they combine PE-Core with Qwen2.5 LM 7B. it outperforms all other models (including InternVL3 which was trained with Qwen2.5LM too!)
The authors release the following checkpoints in sizes base, large and giant:
Authors release following datasets 📑 > PE Video: Gigantic video datasete of 1M videos with 120k expert annotations ⏯️ > PLM-Video and PLM-Image: Human and auto-annotated image and video datasets on region-based tasks > PLM-VideoBench: New video benchmark on MCQA
Dropping the domain-specific downstream image classification content moderation models, including the anime image type classification, GeoSceneNet, indoor-outdoor scene classification, and black-and-white vs. colored image classification models, along with the datasets. 🔥
