Throughout my journey in developing multimodal models, I’ve always wanted a framework that lets me plug & play modality encoders/decoders on top of an auto-regressive LLM. I want to prototype fast, try new architectures, and have my demo files scale effortlessly — with full

🚀 Releasing LMMs Engine by EvolvingLMMs‑Lab — a lean, flexible framework for any-to-any modality pretraining & fine-tuning. 🔧 Built with cutting-edge optimizations: FSDP2, Ulysses Sequence Parallel, Flash Attention 2 📚 Dive in: github.com/EvolvingLMMs-L…

🔥One-Stop Training Engine for Unified Models🔥 ⚡️LMMs-Engine⚡️ is a lean and flexible unified model training engine built for hacking at scale * Support multimodal inputs and outputs, from AR, diffusion and linear models, to unified models like BAGEL 🏠github.com/EvolvingLMMs-L…

🚀 Evaluating MLLMs on Spatial Intelligence is now made EASI! We introduce EASI, an easy-to-use framework and leaderboard for holistic evaluation of multimodal LLMs on spatial intelligence, a key yet underexplored capability. (1/3)

We now have new benchmarks for unified models and we can now use lmms-eval to evaluate BAGEL. github.com/EvolvingLMMs-L…

🔥The giants are entering the arena. 📷 Gemini 3.0 Pro is taking on the Video-MMMU challenge. When the world's smartest models need to prove they master academic video reasoning, they come to our doorstep. The bar has been raised. Who’s next? Explore here:

Scaling Spatial Intelligence with Multimodal Foundation Models

🥳Year-End Reflection on the Growth of LMMs-Lab🥳 2025 has been a fruitful year for 🧠LMMs-Lab🧠 lmms-lab (lmms-lab.com), a non-profit open-source research organization dedicated to feeling and building the future of multimodal intelligence with: 🌟 > 12,000 Total

🚀 EASI v0.2.0 is out! EASI is a unified evaluation suite for Spatial Intelligence, now supporting dual backends: LMMs-Eval and VLMEvalKit for 23 models × 25 spatial benchmarks! 🔗 Code: github.com/EvolvingLMMs-L… 🏆 Leaderboard: huggingface.co/spaces/lmms-la…

我去，太猛了！ 天下苦 Overleaf 久矣，赶 paper 在 chatgpt 和编辑器之间反复横跳真的会疯。prism 出来后又感觉太繁重且不支持中文！ 痛定思痛，我们直接手搓了一个，不管有没有人用，我自己一定会用的产品！ 一个内置 AI Agent 的终极 LaTeX 本地编辑器 LMMs-Lab Writer。 大概有下面几个功能 ✎

Introducing Engram Teams — persistent memory for AI agent swarms. Agent Swarms are the hottest thing in AI right now. OpenAI shipped Swarm. Anthropic just launched Claude Code Agent Teams. Coordination problem? Solved. But these agents have NO MEMORY. They can't learn. They

We are improving lmms-eval to serve for large-scale evaluation. Now, `adaptive mode` gives you 7x improvement over existing baseline, and no boundaries, as long as the model runtimes support. Feel free to leave a comment below if you have any good suggestions, or feedbacks, or

Hi! Yes, we are still improving lmms-eval. Evaluation is important, and we have a lot to do, to make it a right tool, for brewing frontier models. Check out what we did in lmms-eval v0.6, suggestions and feedbacks are welcome.

find the cutest paws across platforms paws.lmms-lab.com

Video might be the next intelligence substrate. Strikingly, video models are beginning to exhibit the same emergent reasoning behaviors first observed in LLMs—multi-path search, self-correction, and layer specialization. We demystify video reasoning and show it doesn’t happen

Memory isn't what users say. It's what they do. How you touch files — and what you change in them — reveals more than anything you'd type into a chat. We built FileGram — agent personalization in a new setting: the File System, where humans and AI actually cowork.