xpeng-robotics

Capek0.5-35B-A3B

README

License: apache-2.0

Capabilities and Training

Capek 0.5 organizes embodied reasoning around four recurring requirements of task execution:

Spatial Reasoning: geometry, relative positions, and viewpoint changes.
Temporal Understanding: event order, scene changes, and temporal localization.
Action Guidance: target points, affordance regions, and ordered image-space trajectories.
State Verification: physical predicates and task completion.

Each capability is first learned by a dedicated specialist through reinforcement learning with verifiable rewards, starting from the same backbone. TIES weight merging combines the specialists to initialize a unified student, which is refined through routed multi-teacher on-policy distillation (MOPD). Specialist routing is used during training; inference runs on a single consolidated checkpoint.

Model Details

Architecture: Qwen3_5MoeForConditionalGeneration (qwen3_5_moe).
Scale: 35B total parameters, approximately 3B active per token, following the upstream architecture.
Inputs: text, images, and video through the supplied multimodal processor and chat template.
Outputs: generated text, including structured predictions when requested by the prompt.

Evaluation Results

Under the paper's matched evaluation protocol, Capek0.5-35B-A3B improves 28 of 34 benchmark rows over its Qwen3.6-35B-A3B initialization. Selected results are shown as Qwen3.6 → Capek 0.5:

VSI-Bench: 60.83 → 70.69; Video-MME: 71.33 → 74.19.
PointBench: 76.66 → 81.22; StateBench-P: 65.80 → 76.80.
EmbodiedBench task success: EB-HAB 46.0% → 63.0%; EB-ALF 50.7% → 55.3%.

License and Attribution

Original Capek 0.5 contributions and the Qwen-derived components are distributed under Apache 2.0, as described in LICENSE.

LICENSE-Apache-2.0: the Apache License, Version 2.0.
LICENSE-Qwen3.6: the unchanged upstream license, including Copyright 2026 Alibaba Cloud.

This is a modified, post-trained derivative. Preserve the applicable license texts, attribution, and modification notices when redistributing. The upstream revision linked in LICENSE identifies the verified license text, not the exact training initialization revision.

Citation

bibtex

@misc{chen2026capek05,  title={Capek 0.5: An Execution-Centric Vision-Language Model for Embodied Intelligence},  author={Ying Chen and Weizhen Li and Zhe Hu and Zhenjiang Li and Rui Jiang and Zhifeng Gu and Lihuang Fang and Jiangping Liu and Lei Yi and Jie Chen},  year={2026},  eprint={2608.06756},  archivePrefix={arXiv},  primaryClass={cs.AI},  url={https://arxiv.org/abs/2608.06756}}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

xpeng-robotics

Model Tree

Base

Qwen/Qwen3.6-35B-A3B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Capabilities and Training

Capek 0.5 organizes embodied reasoning around four recurring requirements of task execution:

Spatial Reasoning: geometry, relative positions, and viewpoint changes.
Temporal Understanding: event order, scene changes, and temporal localization.
Action Guidance: target points, affordance regions, and ordered image-space trajectories.
State Verification: physical predicates and task completion.

Model Details

Architecture: Qwen3_5MoeForConditionalGeneration (qwen3_5_moe).
Scale: 35B total parameters, approximately 3B active per token, following the upstream architecture.
Inputs: text, images, and video through the supplied multimodal processor and chat template.
Outputs: generated text, including structured predictions when requested by the prompt.

Evaluation Results

Under the paper's matched evaluation protocol, Capek0.5-35B-A3B improves 28 of 34 benchmark rows over its Qwen3.6-35B-A3B initialization. Selected results are shown as Qwen3.6 → Capek 0.5:

VSI-Bench: 60.83 → 70.69; Video-MME: 71.33 → 74.19.
PointBench: 76.66 → 81.22; StateBench-P: 65.80 → 76.80.
EmbodiedBench task success: EB-HAB 46.0% → 63.0%; EB-ALF 50.7% → 55.3%.

License and Attribution

Original Capek 0.5 contributions and the Qwen-derived components are distributed under Apache 2.0, as described in LICENSE.

LICENSE-Apache-2.0: the Apache License, Version 2.0.
LICENSE-Qwen3.6: the unchanged upstream license, including Copyright 2026 Alibaba Cloud.

Citation

bibtex

@misc{chen2026capek05,  title={Capek 0.5: An Execution-Centric Vision-Language Model for Embodied Intelligence},  author={Ying Chen and Weizhen Li and Zhe Hu and Zhenjiang Li and Rui Jiang and Zhifeng Gu and Lihuang Fang and Jiangping Liu and Lei Yi and Jie Chen},  year={2026},  eprint={2608.06756},  archivePrefix={arXiv},  primaryClass={cs.AI},  url={https://arxiv.org/abs/2608.06756}}

@misc{chen2026capek05,  title={Capek 0.5: An Execution-Centric Vision-Language Model for Embodied Intelligence},  author={Ying Chen and Weizhen Li and Zhe Hu and Zhenjiang Li and Rui Jiang and Zhifeng Gu and Lihuang Fang and Jiangping Liu and Lei Yi and Jie Chen},  year={2026},  eprint={2608.06756},  archivePrefix={arXiv},  primaryClass={cs.AI},  url={https://arxiv.org/abs/2608.06756}}