GR-3

ByteDance, China · July 2025

ByteDance Seed

Foundation VLAClosed weightsAlso written GR3
Parameters
4 B
The report states that GR-3 contains 4B parameters.
GPU memory
7.5 GB
weights at bf16, computed. Fits a 12 GB card
Inference latency
not published
per action step
Weights
closed
no public checkpoint

What it is

GR-3 is ByteDance Seed's July 2025 VLA and a full departure from GR-2's video-prediction design. It is a 4B parameter model built on Qwen2.5-VL-3B-Instruct with an action diffusion transformer that has half the backbone's layers and reads only the KV cache from the backbone's latter half, an explicit inference-cost decision. The training recipe combines web-scale vision-language co-training, imitation learning on robot trajectories and few-shot fine-tuning from VR-collected human trajectories, with about 450 human trajectories (roughly 30 minutes) sufficient for adaptation. Against pi-0 on ByteDance's ByteMini bi-manual mobile robot it reports 77.1 versus 40 percent on unseen instructions and 97.5 versus 53.8 percent on table-bussing instruction following. Neither code nor weights were released, so for anyone outside ByteDance the transferable value is the data recipe, not the model.

Architecture

Backbone
Qwen2.5-VL-3B-Instruct
Action head
Action diffusion transformer (DiT) with half the layers of the VLM backbone, consuming only the KV cache from the latter half of the VLM layers for efficient inference. Flow-style integration with a step size of 0.2 at inference.
Parameters
The report states that GR-3 contains 4B parameters. The vision-language backbone is Qwen2.5-VL-3B-Instruct and the action diffusion transformer has half the number of layers of the VLM backbone, so roughly 1B sits in the action DiT.
Pretraining data
A three-part recipe: co-training with web-scale vision-language data, fine-tuning from human trajectory data collected via VR devices, and imitation learning on robot trajectories. Concrete task-level figures include 35k robot trajectories (69 hours) for pick-and-place, 101 hours for table bussing, 116 hours for cloth manipulation, and about 450 human VR trajectories (roughly 30 minutes) for the few-shot adaptation experiments.
Embodiments
ByteMini bi-manual mobile robot (introduced alongside GR-3), with 19 controlled DoF excluding the 3 DoF lift mechanism and the head

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 4 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp3215 GBTraining master weights, rarely used for inference
bf16 or fp167.5 GBThe usual way these checkpoints are served
int83.7 GBQuantised, expect some loss of precision on fine motions
int41.9 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 12 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Nothing published. The report describes an architectural efficiency measure (the action DiT reads only the KV cache of the latter half of the VLM layers) but gives no GPU, VRAM or latency number.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

No results in this category are published for this model.

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Real robot (ByteMini), table bussing (long-horizon) instruction following
    97.5%
    success rateAuthors' own evaluation. pi-0 scores 53.8 percent in the same setting.source
  • Real robot (ByteMini), cloth manipulation (deformable, bi-manual) basic setting
    86.7%
    average task progressAuthors' own evaluation. Task progress, not binary success, so not directly comparable with the other rows.source
  • Real robot (ByteMini), pick-and-place unseen instructions
    77.1%
    success rateAuthors' own evaluation. pi-0 scores 40 percent in the same setting.source
  • Real robot (ByteMini), pick-and-place unseen objects
    57.8%
    success rateAuthors' own evaluation. Notably weaker than the unseen-instruction row, which locates the generalization bottleneck on the visual side.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Evaluated on the ByteMini bi-manual mobile robot across pick-and-place, table bussing and cloth manipulation. GR-3 reaches 77.1 percent on unseen instructions against 40 percent for pi-0, 57.8 percent on unseen objects, 97.5 percent on table-bussing instruction following against 53.8 percent for pi-0, and 86.7 percent average task progress on basic cloth manipulation.

Fine tuning it yourself

No public fine-tuning path exists, since neither code nor weights are released. The report's transferable finding is the data recipe: about 450 human VR trajectories, roughly 30 minutes of collection, were enough for meaningful adaptation to new objects and instructions, which is a useful planning figure even without the model.

Where it helps, where it does not

Strengths

  • Strong margin over pi-0 on generalization to unseen instructions (77.1 versus 40 percent) and on instruction following during long-horizon table bussing (97.5 versus 53.8 percent)
  • Efficient action head design: the DiT uses only the KV cache from the latter half of the VLM layers rather than the whole stack
  • Covers bi-manual, deformable-object and mobile manipulation on one platform, which few models in this comparison attempt
  • Quantifies the human-VR data cost for adaptation at about 450 trajectories or 30 minutes
  • Compact 4B total on a 3B Qwen2.5-VL backbone

Limits

  • No code and no weights, so it cannot be run, fine-tuned or independently verified
  • Unseen-object generalization is only 57.8 percent, notably weaker than unseen-instruction generalization at 77.1 percent, which suggests visual rather than linguistic generalization is the bottleneck
  • All results are on ByteDance's own ByteMini robot, a platform nobody outside ByteDance has
  • No published VRAM, latency, control frequency or action chunk size
  • Cloth manipulation is reported as average task progress rather than binary success, so it is not comparable with the success-rate rows

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.