Open Source RoboticsRobots, software & learning

Spirit-v1.5

Robotics model

VLA model implementation with RoboChallenge execution wrapper and separate base/fine-tuned checkpoints. [1] [3]

Setup & usage →CheckpointsModel hub / homepageGetting started

For
Robot-learning researchers familiar with PyTorch and benchmark-specific robot interfaces [1]
Parameters
5,400,901,236 (~5.40B F32 per HF safetensors metadata) [31]
Embodiments
Bimanual 14-DoF manipulation robots (3-camera setup in config.json), RoboChallenge Table30 benchmark robots [1] [31]
Source rights
Some licenses are unresolved. Review licenses

Source researched · Not tested by us
Sources checked · Research details

Licenses & related hardware
License
MIT, Apache-2.0 · license not fully assessed for datasets · Details by part

Requirements · Research details and dates

On this page

Overview

Spirit-v1.5 (Spirit-AI-robotics/Spirit-v1.5) is a 5,400,901,236-parameter (~5.40B) Vision-Language-Action model from the Spirit AI Team pairing a Qwen/Qwen3-VL-4B-Instruct vision-language backbone with a 16-layer Diffusion Transformer (DiT) action head (dit_hidden_size: 1536, 32 attention heads, interleaved self-attention, and cross-attention dim 2560). Both the base model and a task-fine-tuned RoboChallenge checkpoint (Spirit-AI-robotics/Spirit-v1.5-for-RoboChallenge-move-objects-into-box) are published as ungated 20.12 GB FP32 safetensors files on Hugging Face Hub. [1] [30] [31] [32]

Model card summary

Architecture

Qwen/Qwen3-VL-4B-Instruct vision-language backbone paired with a 16-layer Diffusion Transformer (DiT) action head (hidden_size: 1536, 32 heads, 10 denoising steps) using interleaved self-attention and cross-attention (dim 2560).

[30] [31]

Base model

Qwen/Qwen3-VL-4B-Instruct

[30] [31]

Parameter count

5,400,901,236 (~5.40B F32 per HF safetensors metadata)

Exact tensor count verified via Hugging Face API.

[31]

Input modalities
rgbproprioceptionlanguage

Three RGB cameras (cam_high, cam_left_wrist, cam_right_wrist at 3x240x320), 14-D state (padded to 32), and text instruction.

[1] [31]

Output modalities
joint-positionsgripper-command

14-D bimanual action chunks (padded to max_action_dim: 32).

[1] [31]

Action space

Continuous DiT action chunks with chunk_size: 60, n_action_steps: 60, num_steps: 10 denoising steps, and default 14-D action vector padded up to max_action_dim: 32 (MIN_MAX normalization).

[31]

Trained / supported embodiments
  • Bimanual 14-DoF manipulation robots (3-camera setup in config.json)
  • RoboChallenge Table30 benchmark robots

[1] [31]

Training datasets
  • Pretraining corpus not publicly released; fine-tuning example uses RoboChallenge/task_table30_move_objects_into_box

[1]

Interfaces
  • Policy inference API (model/modeling_spirit_vla.py)
  • RoboChallenge HTTP runner (scripts/run_robochallenge.sh)
  • PyTorch FSDP training launcher (scripts/run_finetune.sh)

[1] [3]

Checkpoints

CheckpointFormatSizeAccessPinned revision
Spirit-AI-robotics/Spirit-v1.5safetensors20.12 GBopen9886fa7552de
Spirit-AI-robotics/Spirit-v1.5-for-RoboChallenge-move-objects-into-boxsafetensors20.12 GBopen03ed7624a1a9

Both ungated 20.12 GB FP32 safetensors checkpoints verified via Hugging Face API.

[31] [32]

Is this for you?

Use cases
  • Fine-tuning the Spirit-v1.5 base VLA on custom bimanual/tabletop manipulation datasets via FSDP
  • Reproducing RoboChallenge Table30 benchmark evaluations with task-fine-tuned checkpoints

[1]

Known limitations
  • Base-card Apache2 and code MIT are distinct declarations; backbone, data and task-checkpoint terms remain separate.
  • Training recommendation8xA10080GB is upstream workload advice, not our requirement measurement.
  • No weights downloaded, model run, hardware command or benchmark reproduction.

Editorial scope assessment from the cited implementation/documentation; no execution or compatibility test.

[1] [2] [3] [4] [30] · Editorial assessment

Safety and limitations
  • Repository README (Python 3.10+, torch 2.8.0, transformers 4.57.1, diffusers 0.35.2) and base model card (Python 3.11+, torch 2.9.1, transformers 4.57.3, diffusers 0.36.0) list different pinned dependency versions that must be reconciled before install.
  • Base checkpoint and task-fine-tuned RoboChallenge checkpoint serve different workflows and are not interchangeable.
  • Pretraining dataset is not publicly released; code is MIT while both inspected checkpoints declare Apache-2.0.

[1] [2] [30] [31] [32]

Compute & fine-tuning

Inference compute
  • Documented hardware: NVIDIA A100 80GB (upstream test environment, not a measured minimum)
  • Precision: fp32
  • Edge / consumer capable: No

20.12 GB FP32 safetensors footprint; exact minimum VRAM and latency_ms not measured here.

[1] [31]

Fine-tuning

[1]

Host computer

NVIDIA A100 80GB GPU tested upstream for inference; 8x A100 80GB recommended upstream for FSDP fine-tuning.

[1]

Operating system
  • Linux

PyTorch CUDA / torchrun Linux workflow documented.

[1] · Editorial assessment

Dependencies
  • Repository README Python3.10+ versus base-card recommended Python3.11+
  • Source reports inference testing on NVIDIA A10080GB, not a measured minimum
  • Separate requirements-base.txt and requirements-train.txt; flash-attn needs matching PyTorch/CUDA

[1] [2] [3] [4] [30]

Skills
  • Python
  • PyTorch
  • FSDP / torchrun
  • CUDA / flash-attn

Inferred from documented PyTorch, FSDP, and flash-attn setup.

[1] · Editorial assessment

Setup & usage

Open the upstream getting-started guide [1]

  • Choose inference-only or training dependencies and a named base/fine-tuned checkpoint; follow the selected runtime configuration rather than treating all checkpoints as the same robot policy.
  • Preserve exact source/dependency/model versions and evaluate the selected workflow before connecting physical hardware or reusing data.

Editorial scope assessment from the cited implementation/documentation; no execution or compatibility test.

[1] [2] [3] [4] [30] · Editorial assessment

Upstream-reported benchmarks

Upstream-reported only: Open Source Robotics has not run training, inference, or hardware evaluations for these benchmarks.

BenchmarkUpstream-reported scoreScope & caveats
RoboChallenge Table30Ranked #1 on RoboChallenge Table30 leaderboard as of 2026-01-11 (upstream README claim)Upstream-reported ranking in README.md and Spirit AI blog; not independently verified by OSR.

[1]

Project notes

Action space and checkpoint scope

At the pinned base revision (9886fa7552de), config.json defines three RGB camera inputs (cam_high, cam_left_wrist, cam_right_wrist at 3x240x320), a 14-D proprioceptive state (max_state_dim: 32), and text instructions, outputting continuous DiT action chunks with chunk_size: 60, n_action_steps: 60, num_steps: 10 denoising steps, and default 14-D bimanual actions (max_action_dim: 32, MIN_MAX normalization). Choosing the base checkpoint for FSDP fine-tuning (scripts/run_finetune.sh) is distinct from running the RoboChallenge evaluation wrapper (scripts/run_robochallenge.sh): fine-tuning installs requirements-train.txt on top of requirements-base.txt and requires DATA_ROOT plus PRETRAINED_PATH pointing to a checkpoint directory containing both model.safetensors and config.json, whereas RoboChallenge evaluation uses requirements-base.txt and requires a TASK_NAME defined in robochallenge/runner/task_info.py, ROBOCHALLENGE_JOB_ID, USER_TOKEN, and CKPT_PATH pointing to the directory containing model.safetensors. [1] [31] [32]

Environment discrepancy, compute, and upstream-reported benchmarks

The repository README specifies Python 3.10+ with torch 2.8.0, transformers 4.57.1, and diffusers 0.35.2 (plus optional flash-attn 2.8.3 for training), whereas the base Hugging Face model card recommends Python 3.11+ with torch 2.9.1, transformers 4.57.3, and diffusers 0.36.0; those pinned versions must be reconciled before installation. Upstream reports inference testing on an NVIDIA A100 80 GB and recommends 8x A100 80 GB for FSDP fine-tuning, and states that Spirit-v1.5 ranked #1 on the RoboChallenge Table30 leaderboard as of January 11, 2026. Minimum VRAM, inference latency, and control frequency are not established in the reviewed evidence—the tested A100 80 GB GPU is not a minimum requirement, and the 60-step chunk length alone does not establish execution duration. All benchmark rankings and GPU figures are upstream-reported; Open Source Robotics has not downloaded weights or run training, fine-tuning, inference, or robot evaluations. [1] [30] [31]

What we checked

Source research only. Selected implementation/model text, repository tree and license declarations were inspected. No upstream compilation, dependency installation, simulator, network/USB/device integration, data collection, benchmark or physical robot test was performed. Source files are linked rather than redistributed; external services and paid products are not endorsed.

Not yet established

  • Documented closed-loop control frequency (Hz)

Updates

Upstream last updated
May 29, 2026 [1]
Latest release
Not established
Profile updated
Nov 21, 2026
Sources last checked
Nov 21, 2026
Next review
Feb 19, 2027
Reviewed revision
6d67377 · combined hardware and software repository

Separate license matrix

Distinct robot-facing model contribution with actual implementation/model content and a bounded documented next step. Review covers source artifacts, not operational performance or a complete tested deployment.

LayerAvailability / license
Codeavailable · MIT
Declaration for the scoped source implementation/documentation; dependencies, external services, model/data assets and inherited file-specific terms require separate checks. No blanket redistribution or operational clearance. [1] [2] [3]
Model weightsavailable · Apache-2.0
Both inspected Hugging Face checkpoints (base Spirit-v1.5 and RoboChallenge-move-objects-into-box) are ungated and declare Apache-2.0; weight binaries not downloaded or executed. [30] [31] [32]
Training datasetsavailable · License not fully assessed
Pretraining corpus is not publicly released; upstream fine-tuning guide references public RoboChallenge task datasets whose licenses remain separate. [1]
Documentationavailable · MIT
Declaration for the scoped source implementation/documentation; dependencies, external services, model/data assets and inherited file-specific terms require separate checks. No blanket redistribution or operational clearance. [1] [2] [3]

Separate code, weight, and dataset rights

The spirit-ai-team/spirit-v1.5 codebase is MIT-licensed, while both inspected Hugging Face checkpoints declare Apache-2.0. The pretraining dataset is not publicly released; upstream's fine-tuning guide references public RoboChallenge task datasets (RoboChallenge/task_table30_move_objects_into_box), whose licenses remain separate. [1] [2] [30] [31] [32]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Correction guidance · Discussion is not enabled · Profile format 0.8.0-draft.1

Change history

2026-11-21 — Migrated 0.7.0-draft.5 -> 0.8.0-draft.1 (kind=model); added verified HF API and config.json metadata (S031, S032) while preserving historical source dates.