Skip to page content
SAVRN
Search Plan a project

Open-weight model · Robotics

smolvla_masterji_27E_10000S

by Jooyoung mmporong/smolvla_masterji_27E_10000S

smolvla_masterji_27E_10000S is an open-weight model for robotics from Jooyoung, released under Apache License 2.0. It has 450M parameters. At 16-bit it needs about 1.1 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.

lerobot/smolvlabase를 SO-101 단일팔 픽앤플레이스 데이터셋 Masterji/dataset으로 파인튜닝한 SmolVLA 정책입니다. 이 저장소는 최종 체크포인트(step 10000)입니다.

Parameters450M
Context—
Weights906.7 MB
Licenseapache-2.0
AccessOpen weights
Monthly Downloads—

Runs On

What it takes to serve smolvla_masterji_27E_10000S (450M parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.

PrecisionWeightsMemory neededCheapest setupPer hourAlso fits
16-bit 0.9 GB 1.1 GB 1x MI300X (192 GB)
Vultr
$1.85 1x H100 $1.99 · 1x MI325X $2.00
8-bit 0.5 GB 0.5 GB 1x MI300X (192 GB)
Vultr
$1.85 1x H100 $1.99 · 1x MI325X $2.00
4-bit 0.2 GB 0.3 GB 1x MI300X (192 GB)
Vultr
$1.85 1x H100 $1.99 · 1x MI325X $2.00

Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 18, 2026.

smolvla_masterji_27E_10000S on every accelerator the SAVRN Index prices, at every precision

Model Card

By Jooyoung, published under apache-2.0, revision 49dddb860241.

lerobot/smolvlabase를 SO-101 단일팔 픽앤플레이스 데이터셋 Masterji/dataset으로 파인튜닝한 SmolVLA 정책입니다. 이 저장소는 최종 체크포인트(step 10000)입니다. - 태스크 문장: Pick up the eclipse and place it on the box - 데이터: 30 에피소드 중 0~26번(27개)으로 학습, 27~29번(3개)은 held-out 평가용 - 관측: observation.state(6), 카메라 top·wrist(640×480) → 정책 입력 camera1·camera2로 매핑 - 액션: 6관절 위치(shoulderpan, shoulderlift, elbowflex, wristflex, wristroll, gripper) flow matching 손실이라 노이즈·시간 표본에 따라 값이 흔들립니다. 다른 모델의 손실과 직접 비교할 수 없습니다. eval loss가 가장 낮은 체크포인트와 최종 체크포인트를 같은 held-out 표본·같은 시드로 비교했습니다. 값은 액션 청크의 평균 절대 오차(관절 도, 그리퍼는 0~100 범위)입니다. flow matching의 eval loss와 실제로 샘플링한 액션의 오차는 순위가 다를 수 있어 둘을 함께 적습니다. 관절별 50스텝 평균 오차: shoulderpan.pos 4.39, shoulderlift.pos 8.90…

Read Jooyoung's full model card

lerobot/smolvla_base를 SO-101 단일팔 픽앤플레이스 데이터셋 Masterji/dataset으로 파인튜닝한 SmolVLA 정책입니다. 이 저장소는 최종 체크포인트(step 10000)입니다.

  • 태스크 문장: Pick up the eclipse and place it on the box
  • 데이터: 30 에피소드 중 0~26번(27개)으로 학습, 27~29번(3개)은 held-out 평가용
  • 관측: observation.state(6), 카메라 top·wrist(640×480) → 정책 입력 camera1·camera2로 매핑
  • 액션: 6관절 위치(shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper)

학습 조건

항목 값
기반 모델 lerobot/smolvla_base (전체 450M, 학습 100M — 비전 인코더 고정, action expert 학습)
스텝 / 배치 10000 / 8
학습률 0.0001 (warmup 333, cosine decay 10000 — lerobot 자동 스케일)
chunk_size / n_action_steps 50 / 50
eval·저장 주기 500 스텝
seed 1000
장비 RTX 5050 Laptop 8GB, lerobot 0.6.1

held-out eval loss

flow matching 손실이라 노이즈·시간 표본에 따라 값이 흔들립니다. 다른 모델의 손실과 직접 비교할 수 없습니다.

step eval_loss
500 0.3220
1000 0.2762
1500 0.2240
2000 0.2592
2500 0.2700
3000 0.2380
3500 0.2174 ← 최저
4000 0.2222
4500 0.2254
5000 0.2552
5500 0.2637
6000 0.2790
6500 0.2836
7000 0.2703
7500 0.2836
8000 0.2829
8500 0.3073
9000 0.3040
9500 0.3018
10000 0.3014

체크포인트 비교

eval loss가 가장 낮은 체크포인트와 최종 체크포인트를 같은 held-out 표본·같은 시드로 비교했습니다. 값은 액션 청크의 평균 절대 오차(관절 도, 그리퍼는 0~100 범위)입니다. flow matching의 eval loss와 실제로 샘플링한 액션의 오차는 순위가 다를 수 있어 둘을 함께 적습니다.

체크포인트 eval_loss h1 (seed 0) h1 (seed 1) h50 (seed 0) h50 (seed 1)
step 3500 0.2174 3.20 3.04 6.73 6.80
step 10000 0.3014 2.35 2.22 6.54 6.62
귀무모형(현재 자세 유지) — 3.36 3.36 15.76 15.76

이 체크포인트의 오프라인 평가 (seed 0, held-out 에피소드 [27, 28, 29], 표본 80개)

50스텝 액션 청크의 평균 절대 오차입니다. 단위는 관절 도(그리퍼는 0~100 범위). 귀무모형은 현재 관절 상태를 그대로 유지한다고 가정한 값입니다.

지평 이 체크포인트 귀무모형(현재 자세 유지)
1 2.35 3.36
10 3.44 6.06
30 5.35 11.50
50 6.54 15.76

관절별 50스텝 평균 오차: shoulder_pan.pos 4.39, shoulder_lift.pos 8.90, elbow_flex.pos 9.73, wrist_flex.pos 6.87, wrist_roll.pos 5.16, gripper.pos 4.17

오프라인 오차는 실로봇 성공률이 아닙니다. 실제 성능은 로봇에서 반복 시도로 확인해야 합니다.

추론

녹화 때와 같은 카메라 이름(top, wrist)과 같은 태스크 문장을 써야 합니다. 카메라 이름 매핑과 정규화는 체크포인트의 전처리기에 저장되어 있어 따로 처리하지 않습니다.

lerobot-rollout \
  --strategy.type=base \
  --policy.path=mmporong/smolvla_masterji_27E_10000S \
  --robot.type=so101_follower --robot.port=<팔로워 포트> --robot.id=<보정 id> \
  --robot.cameras="{ top: {type: opencv, index_or_path: <top 카메라>, width: 640, height: 480, fps: 30}, wrist: {type: opencv, index_or_path: <wrist 카메라>, width: 640, height: 480, fps: 30} }" \
  --task="Pick up the eclipse and place it on the box" --duration=60

Identity and Version

Repository
mmporong/smolvla_masterji_27E_10000S
Publisher
Jooyoung
Task
Robotics
Modality
Control
Library
lerobot
Parameters
450M parameters
Languages
Not stated by the source
Revision
49dddb860241060edd77cc89759833aca2c5dde4
First published
2026-10-06
Last updated
2026-10-06

Files and Weights

12 files, 907.5 MB in total. The weights are 3 files totalling 906.7 MB in safetensors.

Weights3 files · 906.7 MB
Configuration6 files · 16.1 KB
Documentation1 file · 4.0 KB
Other1 file · 799.9 KB
Repository1 file · 1.5 KB
Every file
FileTypeSizeSHA-256
model.safetensorsWeights906.7 MB 697aae1843ab
policy_postprocessor_step_0_unnormalizer_processor.safetensorsWeights7.5 KB 7ce00019825e
policy_preprocessor_step_5_normalizer_processor.safetensorsWeights7.5 KB 7ce00019825e
config.jsonConfiguration2.5 KB —
eval/offline_eval_seed0.jsonConfiguration1.8 KB —
eval/offline_eval_seed1.jsonConfiguration1.8 KB —
policy_postprocessor.jsonConfiguration660 B —
policy_preprocessor.jsonConfiguration2.0 KB —
train_config.jsonConfiguration7.3 KB —
README.mdDocumentation4.0 KB —
eval/train.logOther799.9 KB —
.gitattributesRepository1.5 KB —

License and Download

License
apache-2.0
Access
Open weights, no gate
Download size
906.7 MB
Download from Jooyoung

Released by Jooyoung through its official repository on Hugging Face. Read the license.

Built From

Memory Requirements

PrecisionWeights in memory
As published906.7 MB
16-bit0.9 GB
8-bit0.5 GB
4-bit0.2 GB

Weights only, from the published parameter count; the key-value cache and runtime add to this.

Questions About smolvla_masterji_27E_10000S

How much GPU memory does smolvla_masterji_27E_10000S need?

About 1.1 GB at 16-bit and 0.3 GB at 4-bit: the weights (450M parameters) plus a working margin. A long context needs more.

What is the cheapest GPU to run smolvla_masterji_27E_10000S on?

At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.

Can I use smolvla_masterji_27E_10000S commercially?

Yes. smolvla_masterji_27E_10000S is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.

Similar Models

Model · Robotics

smolvla_base

LeRobot

SmolVLA is a compact, efficient Vision-Language-Action (VLA) model designed for affordable robotics, trainable on a single GPU and deployable on consumer hardware, while matching the performance of much larger VLAs through community-driven data. Original paper: (SmolVLA: A Vision-Language-Action Model for Affordable and Efficient Robotics)[https://arxiv.org/abs/2506.01844] For full installation details (including optional video dependencies such as ffmpeg for torchcodec), see the official documentation: https://huggingface.co/docs/lerobot/installation If you’re training / fine-tuning, you typically call forward(...) to get a loss and then: - -policy.chunksize=... - -policy.nactionsteps=...…

Open weights apache-2.0 450M parameters lerobot

Model · Robotics

smolvla_libero

LeRobot

SmolVLA is a compact, efficient vision-language-action model that achieves competitive performance at reduced computational costs and can be deployed on consumer-grade hardware. This policy has been trained and pushed to the Hub using LeRobot. See the full documentation at LeRobot Docs. For a complete walkthrough, see the training guide. Below is the short version on how to train and run inference/eval: Writes checkpoints to outputs/train/ /checkpoints/. Prefix the dataset repo with eval\ and supply --policy.path pointing to a local or hub checkpoint.

Open weights apache-2.0 450M parameters lerobot

Model · Robotics

smolvla_robotwin

LeRobot

SmolVLA is a compact, efficient vision-language-action model that achieves competitive performance at reduced computational costs and can be deployed on consumer-grade hardware. This policy has been trained and pushed to the Hub using LeRobot. See the full documentation at LeRobot Docs. For a complete walkthrough, see the training guide. Below is the short version on how to train and run inference/eval: Writes checkpoints to outputs/train/ /checkpoints/. Prefix the dataset repo with eval\ and supply --policy.path pointing to a local or hub checkpoint.

Open weights apache-2.0 450M parameters lerobot

SmolVLA is a compact, efficient vision-language-action model that achieves competitive performance at reduced computational costs and can be deployed on consumer-grade hardware. This policy has been trained and pushed to the Hub using LeRobot. Learn how to train and run it in the LeRobot smolvla guide, or browse the full documentation. The policy consumes these observation features and produces these action features. Inputs Outputs New to LeRobot? These guides cover the full workflow: - Install LeRobot — set up the lerobot package. - Hardware setup — assemble, wire, and calibrate your robot and cameras. - Record data & train a policy — the end-to-end imitation-learning walkthrough. - CLI…

Open weights apache-2.0 450M parameters lerobot

Model · Robotics

smolvla-policy

Hayato Nakamura

SmolVLA is a compact, efficient vision-language-action model that achieves competitive performance at reduced computational costs and can be deployed on consumer-grade hardware. This policy has been trained and pushed to the Hub using LeRobot. Learn how to train and run it in the LeRobot smolvla guide, or browse the full documentation. The policy consumes these observation features and produces these action features. Inputs Outputs New to LeRobot? These guides cover the full workflow: - Install LeRobot — set up the lerobot package. - Hardware setup — assemble, wire, and calibrate your robot and cameras. - Record data & train a policy — the end-to-end imitation-learning walkthrough. - CLI…

Open weights apache-2.0 450M parameters lerobot

Model · Robotics

thali_smolvla

Prashant Thakur

lerobot/smolvlabase fine-tuned on Prashant-77/thaliall (1050 scripted-expert episodes, 7 skills, language-conditioned, 3 cameras). Camera keys are renamed at train and inference time: overhead → camera1, wrista → camera2, wristb → camera3 (--renamemap; runtime/executors.py applies the same map). Checkpoints in this repo. Root = 20 000 total steps. step14000/ = the best per-skill checkpoint (14 000 steps at batch 16 on a T4; the last 6 000 steps ran at batch 4 with a fresh optimizer on a smaller GPU and lost ground). Policy-only success per skill from task-consistent start states, 20 held-out seeds (eval/skilleval.py --kind smolvla): The scripted expert reaches 9/10 on the full task; the…

Open weights apache-2.0 450M parameters lerobot