Multimodal 🖼️ > ByteDance released SA2VA: a family of vision LMs that can take image, video, text and visual prompts > moondream2 is out with new capabilities like outputting structured data and gaze detection! > Dataset: Alibaba DAMO lab released multimodal textbook — 22k hours worth of samples from instruction videos 🤯 > Dataset: SciCap captioning on scientific documents benchmark dataset is released along with the challenge!
LLMs 💬 > Microsoft released Phi-4, sota open-source 14B language model 🔥 > Dolphin is back with Dolphin 3.0 Llama 3.1 8B 🐬🐬 > Prime-RL released Eurus-2-7B-PRIME a new language model trained using PRIME alignment > SmallThinker-3B is a new small reasoning LM based on Owen2.5-3B-Instruct 💭 > Dataset: QWQ-LONGCOT-500K is the dataset used to train SmallThinker, generated using QwQ-32B-preview 📕 > Dataset: @cfahlgren1 released React Code Instructions: a dataset of code instruction-code pairs 📕 > Dataset: Qwen team is on the roll, they just released CodeElo, a dataset of code preferences 👩🏻💻
Embeddings 🔖 > @MoritzLaurer released zero-shot version of ModernBERT large 👏 > KaLM is a new family of performant multilingual embedding models with MIT license built using Qwen2-0.5B
Image/Video Generation ⏯️ > NVIDIA released Cosmos, a new family of diffusion/autoregressive World Foundation Models generating worlds from images, videos and texts 🔥 > Adobe released TransPixar: a new text-to-video model that can generate assets with transparent backgrounds (a first!) > Dataset: fal released cosmos-openvid-1m Cosmos-tokenized OpenVid-1M with samples from OpenVid-1M
Others > Prior Labs released TabPFNv2, the best tabular transformer is out for classification and regression > Metagene-1 is a new RNA language model that can be used for pathogen detection, zero-shot embedding and genome understanding
> The models are capable of tasks involving vision-language understanding and visual referrals (referring segmentation) both for images and videos ⏯️
> The models come in 1B, 4B and 8B and are based on InternVL2.5 for base architecture and Qwen2, Qwen2.5 and InternLM2 for language model part (depending on the checkpoint)
> The model is very interesting, it has different encoders for different modalities each (visual prompt, text prompt, image and video) then it concatenates these to feed into LLM 💬
the output segmentation tokens are passed to SAM2, to sort of match text (captions or semantic classes) to masks ⤵️
> Their annotation pipeline is also interesting, they seems to use two open large vision LMs to refine the annotations, and have different levels of descriptions to provide consistency.
After 6 years, BERT, the workhorse of encoder models, finally gets a replacement: 𝗪𝗲𝗹𝗰𝗼𝗺𝗲 𝗠𝗼𝗱𝗲𝗿𝗻𝗕𝗘𝗥𝗧! 🤗
We talk a lot about ✨Generative AI✨, meaning "Decoder version of the Transformers architecture", but this is only one of the ways to build LLMs: encoder models, that turn a sentence in a vector, are maybe even more widely used in industry than generative models.
The workhorse for this category has been BERT since its release in 2018 (that's prehistory for LLMs).
It's not a fancy 100B parameters supermodel (just a few hundred millions), but it's an excellent workhorse, kind of a Honda Civic for LLMs.
Many applications use BERT-family models - the top models in this category cumulate millions of downloads on the Hub.
➡️ Now a collaboration between Answer.AI and LightOn just introduced BERT's replacement: ModernBERT.
𝗧𝗟;𝗗𝗥: 🏛️ Architecture changes: ⇒ First, standard modernizations: - Rotary positional embeddings (RoPE) - Replace GeLU with GeGLU, - Use Flash Attention 2 ✨ The team also introduced innovative techniques like alternating attention instead of full attention, and sequence packing to get rid of padding overhead.
🥇 As a result, the model tops the game of encoder models: It beats previous standard DeBERTaV3 for 1/5th the memory footprint, and runs 4x faster!
🕰️ Llama-3.1-405B took 39 million GPU-hours to train, i.e. about 4.5 thousand years.
👴🏻 If they had needed all this time, we would have GPU stories from the time of Pharaoh 𓂀: "Alas, Lord of Two Lands, the shipment of counting-stones arriving from Cathay was lost to pirates, this shall delay the building of your computing temple by many moons "
🛠️ But instead, they just parallelized the training on 24k H100s, which made it take just a few months. This required parallelizing across 4 dimensions: data, tensor, context, pipeline. And it is infamously hard to do, making for bloated code repos that hold together only by magic.
🤏 𝗕𝘂𝘁 𝗻𝗼𝘄 𝘄𝗲 𝗱𝗼𝗻'𝘁 𝗻𝗲𝗲𝗱 𝗵𝘂𝗴𝗲 𝗿𝗲𝗽𝗼𝘀 𝗮𝗻𝘆𝗺𝗼𝗿𝗲! Instead of building mega-training codes, Hugging Face colleagues cooked in the other direction, towards tiny 4D parallelism libs. A team has built Nanotron, already widely used in industry. And now a team releases Picotron, a radical approach to code 4D Parallelism in just a few hundred lines of code, a real engineering prowess, making it much easier to understand what's actually happening!
⚡ 𝗜𝘁'𝘀 𝘁𝗶𝗻𝘆, 𝘆𝗲𝘁 𝗽𝗼𝘄𝗲𝗿𝗳𝘂𝗹: Counting in MFU (Model FLOPs Utilization, how much the model actually uses all the compute potential), this lib reaches ~50% on SmolLM-1.7B model with 8 H100 GPUs, which is really close to what huge libs would reach. (Caution: the team is leading further benchmarks to verify this)
The paper has a lot of experiments (they trained 84 models!) about what makes the video LMs work ⏯️
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 📈 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 🔥 > 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 🔥
Introducing the Synthetic Data Generator, a user-friendly application that takes a no-code approach to creating custom datasets with Large Language Models (LLMs). The best part: A simple step-by-step process, making dataset creation a non-technical breeze, allowing anyone to create datasets and models in minutes and without any code.
Current LLMs process text by first splitting it into tokens. They use a module named "tokenizer", that -spl-it-s- th-e- te-xt- in-to- arbitrary tokens depending on a fixed dictionnary. On the Hub you can find this dictionary in a model's files under tokenizer.json.
➡️ This process is called BPE tokenization. It is suboptimal, everyone says it. It breaks text into predefined chunks that often fail to capture the nuance of language. But it has been a necessary evil in language models since their inception.
💥 In Byte Latent Transformer (BLT), Meta researchers propose an elegant solution by eliminating tokenization entirely, working directly with raw bytes while maintaining efficiency through dynamic "patches."
This had been tried before with different byte-level tokenizations, but it's the first time that an architecture of this type scales as well as BPE tokenization. And it could mean a real paradigm shift! 👏👏
🏗️ 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲: Instead of a lightweight tokenizer, BLT has a lightweight encoder that process raw bytes into patches. Then the patches are processed by the main heavy-duty transformers as we do normally (but for patches of bytes instead of tokens), before converting back to bytes.
🧩 𝗗𝘆𝗻𝗮𝗺𝗶𝗰 𝗣𝗮𝘁𝗰𝗵𝗶𝗻𝗴: Instead of fixed tokens, BLT groups bytes based on their predictability (measured by entropy) - using more compute for complex sequences and efficiently handling simple ones. This allows efficient processing while maintaining byte-level understanding.
I hope this breakthrough is confirmed and we can get rid of all the tokenizer stuff, it will make model handling easier!
💥 𝗚𝗼𝗼𝗴𝗹𝗲 𝗿𝗲𝗹𝗲𝗮𝘀𝗲𝘀 𝗚𝗲𝗺𝗶𝗻𝗶 𝟮.𝟬, 𝘀𝘁𝗮𝗿𝘁𝗶𝗻𝗴 𝘄𝗶𝘁𝗵 𝗮 𝗙𝗹𝗮𝘀𝗵 𝗺𝗼𝗱𝗲𝗹 𝘁𝗵𝗮𝘁 𝘀𝘁𝗲𝗮𝗺𝗿𝗼𝗹𝗹𝘀 𝗚𝗣𝗧-𝟰𝗼 𝗮𝗻𝗱 𝗖𝗹𝗮𝘂𝗱𝗲-𝟯.𝟲 𝗦𝗼𝗻𝗻𝗲𝘁! And they start a huge effort on agentic capabilities.
🚀 The performance improvements are crazy for such a fast model: ‣ Gemini 2.0 Flash outperforms the previous 1.5 Pro model at twice the speed ‣ Now supports both input AND output of images, video, audio and text ‣ Can natively use tools like Google Search and execute code
➡️ If the price is on par with previous Flash iteration ($0.30 / M tokens, to compare with GPT-4o's $1.25) the competition will have a big problem with this 4x cheaper model that gets better benchmarks 🤯
🤖 What about the agentic capabilities?
‣ Project Astra: A universal AI assistant that can use Google Search, Lens and Maps ‣ Project Mariner: A Chrome extension that can complete complex web tasks (83.5% success rate on WebVoyager benchmark, this is really impressive!) ‣ Jules: An AI coding agent that integrates with GitHub workflows
I'll be eagerly awaiting further news from Google!
