Flow - 4x faster than typing
Mini toolsPan-4B
By Pantograph
Pan-4B is Pantograph's largest goal-conditioned Minecraft model, built to test a method for learning goal-directed behavior from internet-scale video. It pretrains on about 500k hours of Minecraft gameplay using hindsight relabeling, learning a state-only value function and next-frame model without action labels, then post-trains on about 2k hours of contractor trajectories with actions to become a controllable policy. At inference it accepts a single goal image plus a 300-frame (about 30 second) video context at 128x128 resolution and 10 FPS, and outputs keyboard and mouse actions at 20Hz. It can walk to targets, build structures, explore for objects, and fight mobs, generalizing to goals never seen during training. It outperforms STEVE-1 and a Gemma 4 based VLA baseline across basics, building, mechanisms, combat, and exploration evaluation categories.
New
Multimodal
Overview
Pan-4B is a 4 billion parameter goal-conditioned agent from Pantograph that plays Minecraft. It is pretrained on about 500,000 hours of internet Minecraft gameplay video using a goal-conditioning method, then post-trained on roughly 2,000 hours of contractor action data. Given a single goal image at inference time, it pursues diverse, previously unseen objectives such as building structures, exploring, and fighting mobs without additional training.
About Pantograph
An applied research lab building generally intelligent, affordable/scalable robots, combining custom low-cost hardware with internet-scale video pretraining.
Industry:
Robotics Engineering
Company Size:
6
Location:
San Francisco, California, US
Website:
pantograph.com
Last updated: July 16, 2026
