---
title: "GigaBrain-0"
description: "GigaBrain-0 from GigaAI: vla and vla action prediction; Verified; 2 variants."
modelId: "giga-brain-0"
pageSource: authored
---
GigaBrain-0 is GigaAI's 3.5B VLA foundation policy with subtask prediction, released as GigaBrain-0-3.5B-Base and the 0.1 refresh under Apache-2.0. In this catalog the in-tree GigaBrain-0.1 runtime emitted a checkpoint-backed CUDA action_trace using converted LeRobot RoboTwin statistics, with PaliGemma tokenizer and the physical-intelligence FAST tokenizer staged.
The **0.1** variant is the promoted GPU route: a public pipeline wrote a checkpoint-backed CUDA `action_trace` on an **80 GB** card using **converted LeRobot RoboTwin** statistics (about 20.8 GB allocated, ~55 s on the recorded probe). That is **non-leaderboard** runtime readiness. Official task-score parity still needs gated official task data and task-specific `norm_stats` / `delta_mask`. The 3.5B-base line passed a tensor GPU probe; promoted action-trace evidence currently covers **0.1**.
The recorded paper is [arXiv 2608.15875](https://arxiv.org/abs/2608.15875).
Project page: [gigaai-research.github.io/GigaBrain-Challenge-2026](https://gigaai-research.github.io/GigaBrain-Challenge-2026/). Code: [open-gigaai/giga-brain-0](https://github.com/open-gigaai/giga-brain-0).
Paper: [arXiv 2608.15875](https://arxiv.org/abs/2608.15875). Project: [gigaai-research.github.io/GigaBrain-Challenge-2026](https://gigaai-research.github.io/GigaBrain-Challenge-2026/). Code: [open-gigaai/giga-brain-0](https://github.com/open-gigaai/giga-brain-0). Also required: gated `google/paligemma-3b-pt-224` (Hugging Face auth) and `physical-intelligence/fast`. Runner: **verified** for that converted-stats scope. Aliases include `giga-brain-0.1` and `GigaBrain-0.1-3.5B-Base`.
## What it is for
Use `giga-brain-0-3.5b-base` for vision-language-action inference on this card. The recipe exists so you can launch the recorded policy profile.
A 3.5B policy on a PaliGemma tokenizer base using the FAST action tokenizer for discrete action encoding, plus subtask prediction heads. Architecture internals beyond the official repository pointer are not recorded in this repository.
Catalog tasks include vla, vla action prediction, vla policy rollout, and robot policy. Catalog aliases include `giga_brain_0`, `gigabrain0`, `giga-brain-0.1`, and `GigaBrain-0-3.5B-Base` among others. Also recorded: `giga-brain-0.1-3.5b-base`.
Recorded evaluation entry points include [RoboTwin 2.0](/docs/evaluation/benchmark-hub/robotwin) (CUDA action trace was produced with converted RoboTwin statistics).
## Runtime contract
Optional `prompt`, `input_path`, `interactions`, and `fps`; artifacts `action_trace` and `giga_brain_0_action_trace.json`.
CLI task profiles: `interactive-video`. WorldFoundry binds `GigaBrain0Pipeline`.
Start on `giga-brain-0-3.5b-base`. The family registers 2 catalog variants (cards above): `giga-brain-0-3.5b-base` (`vla.action_prediction`) — Runner parity pending; `giga-brain-0.1-3.5b-base` (`vla.policy_rollout`) — Runner parity pending. Family status: **Verified**.
The PaliGemma tokenizer model is gated (manual review) and requires Hugging Face auth. Official task-score parity remains separate because official task norm_stats/delta_mask assets are gated or not staged.
WorldFoundry in-tree GigaBrain-0.1 runtime emitted a checkpoint-backed CUDA action_trace using converted LeRobot RobotWin stats.
## Running in WorldFoundry
Launch `worldfoundry-eval run giga-brain-0-3.5b-base --pipeline.task-profile interactive-video` to execute the policy. The command builder above switches `--pipeline.task-profile` among recorded tasks and still emits `worldfoundry-eval run`.
It runs in the unified environment `worldfoundry-unified-cu128` (Python 3.11, CUDA 12.8).
Recorded weight repositories include `open-gigaai/GigaBrain-0-3.5B-Base`, `open-gigaai/GigaBrain-0.1-3.5B-Base`, and `google/paligemma-3b-pt-224` among others. Prepare the environment and check local assets with `bash scripts/inference/prepare_model_infer.sh giga-brain-0-3.5b-base`. Add `--download` to fetch missing public checkpoints.
Recorded licenses include apache-2.0; status: **Verified**.