DeepCybo · 2026-09-14 · major
PhysBrain 1.5 report — an open 8B model tops 28 embodied AI tests
PhysBrain 1.5 is DeepCybo's open 8B model that understands a scene, generates robot motion and predicts the next state in one network. It averages 72.5 across 28 embodied benchmarks, the best open-source result reported.
One open 8B network reads a scene, plans the arm's next move, and predicts what the scene looks like a second later.
Quick facts
| Maker | DeepCybo, with Zhongguancun Academy |
|---|---|
| Sizes | 2B and 8B checkpoints |
| Backbone | Qwen3-VL |
| Benchmark average | 72.5 across 28 embodied tests |
| Best open-source on | 14 of the 28 benchmarks |
| Availability | Open weights on Hugging Face plus a demo Space |
| Paper | arXiv 2609.14973, submitted 14 Sep 2026 |
Benchmarks
| GPT-6-Astra | 73.3 pts | |
|---|---|---|
| Gemini 3.6 Flash | 73 pts | |
| PhysBrain 1.5-8B | 72.5 pts |
What is it?
PhysBrain 1.5 folds three robotics jobs into a single open model: understanding what a scene shows, generating the motion to act on it, and predicting how that scene evolves. DeepCybo published the technical report on arXiv on 14 September 2026, documenting the recipe behind checkpoints already posted on Hugging Face in 2B and 8B sizes. The work is credited to the DeepCybo team with Zhongguancun Academy.
How does it work?
Starting from a Qwen3-VL vision-language backbone, the team encodes language replies, end-effector motion and dense visual targets as discrete tokens over one shared vocabulary, then trains all three with ordinary next-token prediction. Pre-training draws its embodied supervision entirely from human interaction videos, using task-centered episodes to pair semantic and spatial context with recovered motion. A later fine-tuning stage mixes human demonstrations, robot trajectories and simulated experience.
Why does it matter?
Robotics teams normally stitch together a separate perception model, a control policy and a world model, each with its own training loop. Averaging 72.5 across 28 embodied understanding benchmarks puts the 8B model a point behind GPT-6-Astra at 73.3 and Gemini 3.6 Flash at 73.0 while beating every other open model — so a lab can download weights instead of renting a closed API for the same work.
Who is it for?
robotics and embodied-AI researchers
Frequently asked questions
- Can I download the PhysBrain 1.5 weights?
- DeepCybo has published PhysBrain 1.5 checkpoints on Hugging Face in both 2B and 8B sizes, in BF16 safetensors format, alongside a hosted demo Space for the 8B model. The report's evaluation harness is released separately as PhysBrainEvalKit on GitHub, so the 28-benchmark average can be reproduced rather than taken on trust.
- How does PhysBrain 1.5 compare to GPT-6-Astra?
- Across 28 embodied understanding benchmarks, PhysBrain 1.5-8B averages 72.5 against 73.3 for GPT-6-Astra and 73.0 for Gemini 3.6 Flash — under a point behind both closed models. PhysBrain 1.5 takes the best open-source score on 14 of those 28 benchmarks and second place on 10, while keeping its general multimodal ability.
- What does PhysBrain 1.5 actually output for a robot?
- Beyond language answers, PhysBrain 1.5 emits structured spatial coordinates, end-effector trajectories, and predicted future scenes as spatially aligned RGB, depth and robot-mask outputs. The report shows these as qualitative examples rather than a hardware trial, so it demonstrates the model can produce robot-usable signals, not that it has been validated on a specific arm.
- What do I need to run PhysBrain 1.5?
- PhysBrain 1.5 integrates with standard infrastructure: Transformers, vLLM and SGLang for inference, and LLaMA-Factory for fine-tuning. The 8B checkpoint is roughly 9B parameters once the extended action and visual-state vocabulary is counted; teams with less GPU memory can start from the 2B checkpoint, which the report excludes from its open-source rankings.
Try it
Model id DeepCybo/PhysBrain1.5-8B on Hugging Face — runs under Transformers, vLLM or SGLang