We create world-class foundation models that redefine what AI can do.
Sunnyvale • Abu Dhabi • Paris
Joined March 2026
- Tweets237
- Following10
- Followers5.3K
- Likes66
Pinned Tweet
Introducing PAN 2: an open foundation world model for physical intelligence.
A language model predicts the next word. PAN 2 predicts the next world: what happens after each action, grounded in time, physics and space.
Towards AI that simulates “what if?” before it acts.
More details: ifm.ai/pan/
PAN 2 is the best open model in physics and robotics.
🥇 #1 open model on Physics-IQ Verified
🥇 #1 open models in our RBench evaluation
Open weights, training code and inference code under Apache 2.0: huggingface.co/IFM/PAN2
Thank you @haoailab for day-0 support of PAN 2 in FastVideo. With FastVideo, PAN 2 simulates 20 seconds in about 2 minutes on 8×B300, so developers can run it fast from the first day.
Try it now: huggingface.co/IFM/PAN2
FastVideo team is excited to announce day-0 support for @IFM_AI PAN 2! 🚀
PAN 2 is a foundation world model for physical intelligence, enabling simulative reasoning, where an agent imagines the outcomes of its candidate actions before choosing among them.
Simulate 20 seconds in 2 minutes on 8×B300. 10 seconds in 44 s. 📷
FastVideo: github.com/hao-ai-lab/FastVi…
PAN 2 Recipe: haoailab.com/FastVideo/cookb…
Congrats to the PAN 2 team at IFM!
@mbzuai
Huge thanks to @vllm_project for day-0 support of PAN 2.
Serve it on your own machine today:
huggingface.co/IFM/PAN2
🎉 Great work from the PAN 2 team on an open-weight world model built for real-world simulation, grounded in time, physics and space to predict how scenes evolve as actions unfold.
vLLM-Omni brings Day-0 support for PAN 2 text-to-video and image-to-video generation, making it available through an open inference stack.
Jais Chat mobile app is now live!🎙
The most capable Arabic AI model is in your pocket.
Skip the typing, talk to your phone in dialects and get the up-to-date information from web searches.
“Jais Chat” free on iOS and Android:
apps.apple.com/ae/app/jais-c…
play.google.com/store/apps/d…
14
10
1,134
In September, we said that K2 Horizon was the largest LLM open drop in AI history yet. We’ve now delivered on the commitment.
When people say open models, they usually mean open weights. K2 Horizon 360 is way more: 8 artifact categories and 30+ repositories across every stage of LLM training.
- Checkpoints: 380 across the six models, pre-training through RL, so you can trace when a capability first appeared
- Data: 13 trillion pre-training tokens across 5 datasets and 24 subsets
- Data tools: 11 repositories, 7 task generation and 3 advanced generation pipelines
- New architecture: MoVA in our 36B-A4B model, MoE sparsity applied to the value vectors in multi-head attention, compatible with efficient attention algorithms at no additional KV-cache cost
- Pre-training infra: xLLM, faster pre-training and fine-tuning, with rework and reconfiguration at lower cost
- Post-training infra: RL360 for long-horizon agentic RL, 8192 concurrent sandboxes, 150 steps in 7 days, ~1.3M rollouts. It separates system crashes from the model’s own mistakes
- Training logs: 58 runs, with loss curves, gradient norms and the configs that actually ran, so you can check your own against a real one at this scale
- Diffusion LLM: Uno, autoregressive quality at diffusion speed
Technical details in previous threads, full artifact directory →
ifm.ai/k2/artifacts/
K2 Horizon now runs on llama.cpp and Ollama, from 0.9B to MoVA-36B.
It runs entirely on local hardware, with no cloud API needed.
GGUF weights: huggingface.co/collections/I…
Ollama: ollama.com/ifm/k2-horizon
Want to run K2 Horizon on your laptop? Start with 0.9B, 3.7B or 7B.
Easy setup with our skill! Hand it to your AI coding agent and ask for a K2 Horizon deployment:
- Picks the size for your machine
- Installs llama.cpp and pulls the official weights
- Leaves a local OpenAI-compatible API running
Download the skill, or copy its link and give it to your AI agent: github.com/MBZUAI-IFM/k2-loc…
Today we dropped a technical report on Jais 3, a six-fleet Arabic-centric agentic models from 0.9B to 375B.
Quick highlights:
- Adapted from K2 Horizon: +9 to +35 on the Arabic average over the base model
- Agentic at small sizes: 3.7B and 7B beat every open model of their size on all 10 agentic benchmarks
- FTPO: a short preference stage after SFT, repetition loops from 13.8% to 1.2%
- Audited Arabic benchmarks: 15 widely used Arabic benchmarks turned out to contain wrong answers, and we audited to remove the 5.3% of items our three judges flagged
Weights, checkpoints from every stage, data samples, logs and eval code, all open 👉
ifm.ai/jais/research
The full training logs for all six models are open, so you can follow each run from the first update to the last. We’re also releasing samples from the 1.63T-token data inventory those runs drew on, including the Arabic dialect and Arabizi corpora we built for Jais 3.
Training logs: forge.coreweave.com/wandb/mb…
Training data samples: huggingface.co/datasets/IFM/…
1
2
407
The most detailed open LLM data release yet: K2 Horizon's pre-training data and post-training data-generation pipelines are now open!
13 trillion pre-training tokens across 5 datasets and 24 subsets. Over 10 trillion tokens of synthetic data spanning mathematics, reasoning, coding, planning, and problem-solving.
Beyond the scale, four highlights for model builders (details in the threads):
- Scaling synthetic data without losing diversity. Our generation methods produce diverse training examples at scale, nearly matching natural web text in measured diversity.
- More verification isn't always better. Strict verification can throw away useful training signal.
- Small-scale wins can mislead. Our chosen data mix actually lost in smaller experiments.
- Beyond teacher imitation. Our post-training data generation methods produce training trajectories the teacher couldn't generate on its own.
Open pre-training datasets, post-training data-generation pipelines, and processing tools for researchers to inspect, reuse, and build on.
Pre-training data: huggingface.co/collections/I…
Post-training data-generation code & tools: github.com/ifm-ai/DATA_NAVIG…
Synthetic data lets us deliberately shape what models learn, rather than relying only on what's available on the web.
At scale, synthetic examples can become repetitive. We use retrieval-augmented generation and diversity knobs to cover a wider range of problems, contexts, and behaviors.
We use compression to quantify diversity and show that our synthetic data is nearly as diverse as natural web text.
We're open-sourcing RL360, the RL infrastructure behind K2 Horizon's six domain experts: 8,192 concurrent sandboxes, 150 steps in 7 days, ~1.3M rollouts.
Long-horizon agentic RL runs thousands of agents in real sandboxes on tasks that take minutes. Some of those sandboxes crash mid-task, and a crash arrives as the same low score as a wrong answer.
Without infrastructure that can keep up, GPUs are underutilized and the model gets penalized for system crashes rather than its own mistakes.
RL360 is built for this. We share how it works, what broke at scale, and what we learned in the thread.
Code: github.com/ifm-ai/RL360
Run: wandb.ai/mbzuai-llm/rl360-pu…
RL360 is modular by design. Rollouts, sandboxes, inference, and training sit behind explicit service boundaries and scale independently: Miles for rollouts and optimization, Harbor for agents and sandboxes, SMG + SGLang for inference, Megatron-LM for training.
The modular design paid off as we scaled to 8,192 rollouts per step. Each time a different layer saturated, we widened only that one.
1
5
1,310
Jais-ASR ranks #1 on the Open Universal Arabic ASR Leaderboard.
Our open model is built for Modern Standard Arabic and dialects from 19 countries.
Beyond the leaderboard, it handles:
- Natural code-switching with English, French, Spanish, and Italian
- Robustness to background noise
- Faithful transcription, resistant to hallucinations
- Arabic diacritization
- Built-in language and dialect identification
Weights released under Apache 2.0: huggingface.co/IFM/Jais-ASR
Jais 3 is built for Arabic from the start.
Speech: the most accurate open Arabic ASR, across MSA and dialects
Writing: fluent, native Arabic from the first line to the last
Agents: leads open models on Arabic agentic tasks
Download weights: huggingface.co/collections/I…
Day-0 support for Jais 3 on @sgl_project, so an Arabic-first agentic model is now straightforward to serve at scale. Thank you to the SGLang team!
Weights: huggingface.co/collections/I…
API: platform.ifm.ai
Jais 3 from @IFM_AI is live on SGLang from Day 0!
Jais 3 is a family of six open Arabic-centric models, from 0.9B to 375B, built to plan, call tools, and work on Arabic documents as agents, ready to serve on SGLang!
Huge thanks to the vLLM team!
With Day-0 support on @vLLM, Jais 3 now runs on both of the serving stacks most teams already have.
Weights: huggingface.co/collections/I…
API: platform.ifm.ai/
📷 Day-0 support for Jais 3 in vLLM!
Congrats to @IFM_AI and collaborators on the launch. Arabic-native agentic AI, ready to serve with vLLM from day one.
Jais 3: First adaption of K2 Horizon
Jais 3 is built on K2 Horizon, our six-fleet fully open model family.
Stronger in Arabic at every size, without starting from scratch. Plus Jais Speech, the most accurate open Arabic ASR.
Download weights now: huggingface.co/collections/I…
Introducing Jais 3, Arabic-native agentic AI.
Jais 3 375B matches the best open models, and with Jais Speech offers the most advanced Arabic AI stack available.
→ Hugging Face: huggingface.co/collections/I…
→ Technical blog (Jais 3): ifm.ai/blog/jais3
→ Technical blog (Jais Speech): ifm.ai/blog/jais-asr
→ Learn more: ifm.ai/jais/
Jais 3 was created by the Institute of Foundation Models (IFM) at MBZUAI, in collaboration with Inception42 and Cerebras Systems.
#Jais3 #IFM #MBZUAI
Jais 3, the first adaptation of K2 Horizon. Coming soon.
#Jais3
Hector Liu, IFM's Head of Technology and US Lab Director, walked MBZUAI through how K2 Horizon was built. A rare, candid close-up of LLM training.
From the room:
- Two new attention designs: softplus gated attention is a more expressive form of gated attention; Mixture-of-Value Attention adds capacity at no extra KV cache cost.
- Nearly half the pre-training data was synthetic. An interesting generation method: Socratic dialogues, a teacher model questioning a student model toward the answer.
- A scaling law guided learning-rate choices across model sizes. Four models trained on the same data learned at the same pace, with normalized loss curves lining up almost perfectly.
More in the upcoming paper.
K2-Type, our open Jev-style decision model, now runs free in your browser on ZeroGPU.
Huge thanks to @multimodalart from @huggingface for spinning up the Space! Ask Yes/No, Choice, or Score questions and get structured answers in one pass, no GPU or API key needed.
Try it 👇
huggingface.co/spaces/IFM/k2…
