PAN 2 and Liquid AI’s LFM2.5-2.6B-Base are both research checkpoints, but they ask very different things of anyone hoping to adapt them. OrcaRouter’s comparison finds that the useful choice is less about which model wins and more about what you are trying to build, and what hardware and data you can bring to it.
Watch Desk analysis
What happened
OrcaRouter compares PAN 2, a 29-billion-parameter video world model from MBZUAI’s Institute of Foundation Models, with Liquid AI’s 2.69-billion-parameter language-model checkpoint. Neither is a finished chat product: LFM2.5-2.6B-Base is intended for further training, while PAN 2 generates video to simulate the consequences of actions.
The comparison says PAN 2’s repository includes training configurations, but adapting it means preparing video latents and structured captions. Its own weights use Apache-2.0, while the third-party components have separate terms. Liquid’s LFM1.0 licence permits commercial use subject to a revenue threshold, so teams need to check which terms fit their plans.
Why it matters
The resource demands are worlds apart. OrcaRouter describes LFM2.5-2.6B-Base as suitable for fine-tuning on a single GPU, while PAN 2’s video-generation setup calls for multi-GPU infrastructure. Its cited recipe uses four B300 GPUs to generate about 13 seconds of video in roughly 119 seconds; the repository’s default setup uses eight H200s.
That makes “open weights” a starting point, not a promise that a model is easy to run or adapt. Data preparation, compute and licence terms all shape whether a checkpoint is useful for a real project.
Our read
This is a practical comparison, not a contest between substitutes. If you want to adapt a compact text model, LFM2.5-2.6B-Base is the more approachable starting point. PAN 2 is for teams with video data, specialist infrastructure and a reason to explore world simulation. The word “open” does not make the GPU bill disappear. OrcaRouter also notes that PAN 2’s evaluation evidence is limited, so teams should treat it as a research artefact rather than a proven simulator.
What to watch
- Whether MBZUAI publishes independent evaluations or a more accessible fine-tuning recipe for PAN 2.
- Whether Liquid AI’s base checkpoint proves useful for domain-specific fine-tuning.
- Whether teams can meet each model’s licence conditions and compute requirements in practice.
Discussion spark: When choosing an open model to adapt, would you prioritise a smaller checkpoint you can fine-tune on one GPU, or a more ambitious research model that demands a multi-GPU setup?
Sources and evidence
- PAN 2 vs LFM2.5-2.6B-Base: Checkpoints, Not Products – OrcaRouter (11 November 2026, 01:45 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.