Guangyi Liu

🚀 Excited to announce the release of PAN, a general world model I’ve been working on for years. PAN can simulate physical, agentic, and nested worlds — generating infinite interactive experiences to train and evaluate AI agents. Check out demo: ifm.mbzuai.ac.ae/pan/ 👇

To mark the 2nd anniversary of LLM360, we are proud to release K2-V2: a 70B reasoning-centric foundation model that delivers frontier capabilities. As a push for "360-open" transparency, we are releasing not only weights, but the full recipe: data composition, training code, logs, and intermediate checkpoints. About K2-V2: 🧠 70B params, reasoning-optimized 🧊 512K context window 🔓 "360-Open" (Data, Logs, Checkpoints) 📈 SOTA on olympiad math and complex logic puzzles

Today, we are releasing a new version of K2 (K2-V2), a 360-open LLM built from scratch as a superior base for reasoning adaptation, while still excelling at core LLM capabilities like conversation, knowledge retrieval, and long-context understanding. K2 fills a major gap: highly capable models with no transparency. Instead of releasing only weights, we’re sharing the full training story — dataset recipes, mid-training checkpoints, logs, code, and evaluation tools. That’s 360-open. What’s inside: • 70B dense transformer engineered as a reasoning-enhanced base model • Native 512K context (extendable via RoPE scaling) • Mid-training reasoning phase • Strong tool-use scaffolding What we’re open-sourcing: • 250M+ reasoning traces (math, planning, multi-step logic) • Full pre- & mid-training data compositions • All mid-training checkpoints • Training logs, code, Eval360 Performance: • GPQA-Diamond: 55.1% mid-training → 69.3% after SFT (strongest fully open 70B model) • KK-8 Logic Puzzles: 83% — competitive with DeepSeek-R1 & OpenAI o3-mini-high • ArenaHard V2: 62.1% — close to Qwen3 235B • Outperforms Qwen2.5-72B and approaches Qwen3-235B despite being smaller and fully transparent. 🔗 The Model: bit.ly/3KIYwuo 🔗Technical Report: bit.ly/49V8h2U 🔗Blog: bit.ly/49V7gb6

@HuggingPapers Check here 🚀x.com/guangyi_l/stat…

MBZUAI's PAN Team unveils PAN: a new world model This general, interactable, and long-horizon world model predicts future states through high-quality video simulation, conditioned on history and natural language actions.

Also check demos here x.com/guangyi_l/stat… and x.com/guangyi_l/stat…

🚀 Excited to announce the release of PAN, a general world model I’ve been working on for years. PAN can simulate physical, agentic, and nested worlds — generating infinite interactive experiences to train and evaluate AI agents. Check out demo: ifm.mbzuai.ac.ae/pan/ 👇

📢 Check out our new PAN World Model with strong performance across the core dimensions that define a modern world model, including simulative reasoning, action simulation and long horizon prediction. PAN builds a generative latent world that brings perception, state, action and causality into one coherent system, and it delivers clear gains over VJEPA2 and other frontier models: x.com/huskydogewoof/… The work also connects nicely to the broader conversation about decoder free joint embedding approaches versus generative decoder approaches. @ericxing offered an insightful perspective on this topic linkedin.com/posts/eric-xin… Here are some great demos from PAN: x.com/guangyi_l/stat…

@gitansgarg Sure

@guangyi_l This is amazing. I am also working on world models and model based RL (specifically JEPA). can i dm you?

🌀 Every choice changes the world. PAN shows how — revealing the chain of cause and effect behind every action. From one world, infinite futures unfold. 🎬 Dive into more demos 👉 ifm.mbzuai.ac.ae/pan/

@guangyi_l incredible work!

@CottenIO Stay tuned

@guangyi_l Code?

@guangyi_l This will be viral in one week when AI influencers pick it up. This is incredible progress

🚀 PAN doesn’t just generate videos — it simulates the world itself. 🌎 Interact, explore, and shape the future you build. 👉 See more: ifm.mbzuai.ac.ae/pan/

🌍 Introducing long-horizon, interactable PAN World Model that lets you step into coherent worlds and evolve them through language-guided actions. A world model is far more than visuals — it understands world dynamics, enabling an agent to imagine scenarios, anticipate outcomes, and act with strategy.

🚀 Introducing PAN, our latest general world model. 💡 Compared to traditional video generation models like Sora 2, PAN simulates worlds you can interact with, over long horizons, with natural-language actions.

PAN represents a major step toward general world modeling — a foundation for scalable, embodied, and interactive intelligence. More details here ifm.mbzuai.ac.ae/pan/ Technical Report: arxiv.org/pdf/2511.09057

🧩 Architecture: PAN is an autoregressive, generative latent predictor — continuously modeling the dynamics of worlds and agents. This enables rich, consistent, open-ended simulation across both physical and cognitive domains.

🏆 Performance: PAN significantly outperforms JEPA-2, Cosmos-2, and other prior world models in: simulative reasoning, action prediction and long-horizon planning.

At its core, PAN connects language, perception, action, and latent thoughts, enabling long-horizon reasoning and simulation. It builds on top of pretrained LLMs and video diffusion models, unifying them into a single model of the world.