Published 9 September 2026
DeepCybo published PhysBrain 1.5-8B and PhysBrain 1.5-2B on 8 September 2026, two models that look at camera images and reason about the physical world so a robot can act on it. The 8B version scores 72.5 on DeepCybo’s own suite of 28 tests, ahead of every other open model it measured and within a point of GPT 6 Astra at 73.3. Neither model carries a licence.
What does PhysBrain 1.5 do?
A robot needs three things from a model. It has to understand what it is looking at, decide what to move next, and have some idea of what the world will look like after it moves. Most systems bolt a separate module onto a shared vision-language model for each of those jobs.
PhysBrain 1.5 does all three with one model and no extra modules. DeepCybo turned robot arm movements and future camera views into new words in the model’s vocabulary. The model then just predicts the next word, whether that word is English, a coordinate for the gripper, or a piece of the image it expects to see one second from now. DeepCybo calls this a closed physical loop: look, act, look again.
The practical payoff is that the file you download stays an ordinary Qwen3-VL model. It runs on the standard serving tools without custom code. The cost, which DeepCybo does not discuss, is that chopping a smooth arm movement into a fixed set of steps loses precision that a purpose built control module would keep.
- PhysBrain 1.5-8B: 8.9 billion parameters, 17.8 GB of weights, built on Qwen3-VL-8B-Instruct
- PhysBrain 1.5-2B: 2.2 billion parameters, 4.3 GB of weights, built on Qwen3-VL-2B-Instruct
- Context window: 262,144 tokens on both sizes
- The only structural change from Qwen3-VL is 16,640 extra vocabulary entries for actions and future views
- Licence: none declared, so you can download the weights with no written permission to use them
PhysBrain 1.5 benchmarks against GPT 6 Astra and Qwen3-VL
The cleanest number in the release is the gap to its own starting point. PhysBrain 1.5-8B was trained on top of Qwen3-VL-8B-Instruct, so the difference between the two columns is what that training actually bought.
| What is measured | PhysBrain 1.5-8B | Qwen3-VL-8B-Instruct |
|---|---|---|
| Average across all 28 benchmarks | 72.5 | 59.5 |
| VLABench (planning a robot task) | 76.4 | 44.8 |
| EgoPlan-Bench2 (planning from a first person view) | 62.1 | 32.1 |
| BLINK (basic visual perception) | 87.9 | 81.7 |
| MMSI-Bench (spatial reasoning across images) | 41.0 | 30.8 |
| RefSpatial-Bench (pointing at the right spot) | 50.9 | 43.6 |
Every one of these figures was produced by DeepCybo. The model card says the team re-ran all comparison models itself so the metrics would match, which is a defensible choice and also means no number here has been checked by anyone else. Against closed models the 8B scores 72.5 to 73.3 for GPT 6 Astra, 73.0 for Gemini 3.6 Flash and 67.9 for Claude Opus 5. The 2B version reaches 66.6, which would beat every other open model in the table, though DeepCybo leaves it out of its own open-source ranking.
What DeepCybo is not saying
There is no licence field on either model card and no licence file in either repository. Anyone can download 17.8 GB of weights and has no written permission to use them. For a company release aimed at robotics teams, that is the first thing to fix.
The model card cites a PhysBrain 1.5 Technical Report and links to a GitHub repository. That link returns a 404, and no paper for PhysBrain 1.5 exists on arXiv. Details of the evaluation are deferred to a report that has not been written.
All 28 benchmarks test one thing: whether the model understands what it is looking at. That is one of the three capabilities in the pitch. Deciding what to move and predicting what comes next are supported only by example pictures, labelled as qualitative results. There are no robot success rates at all, in simulation or on hardware, which is a step back from PhysBrain 1.0, where Chinese press reported an average success rate of 80.2% in simulation.
The three closed models were also handicapped. GPT 6 Astra was run at a low thinking setting and Gemini 3.6 Flash at minimal, while two of the open competitors were run with thinking switched on. Coming within 0.8 points of a model that was told not to think hard is not the same as matching it.
What this means
This is worth testing if you are a robotics team that already runs a vision-language model as the reasoning layer above a separate control policy, and you want to know whether a single checkpoint can replace the stack. The 13 point jump over the Qwen3-VL model it was trained on is real and large, it runs on standard serving tools, and the 2B version fits on a consumer graphics card, which makes it one of the few models of this kind you can put on the robot itself rather than in a data centre. Download it, run your own perception tests, and treat the ranking claim as a starting hypothesis.
Do not put it anywhere near production yet. A model with no licence, no technical report, a citation pointing at a dead URL and zero published control results is a research preview that has been labelled as a release. The honest read is that DeepCybo has shown a strong perception model and asserted a robotics model on top of it. Watch for two things: a licence, and any number at all showing a robot completing a task. Until the second one arrives, the interesting claim in this release remains untested.
For more information, visit the official announcement of PhysBrain 1.5 in the PhysBrain 1.5-8B model card
