ATQ label-gated MoE, 2.5x compressed group, v22 VLM-only conf label, SO(3)-composed merged rotation GT, RoboCasa 60k 60,000 steps, seed 42, RoboCasa 24 tasks, 2 GPUs.
Runs On
What it takes to serve atq-v22so3-labelgated-2.5x-60k (2.8B parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 5.7 GB | 6.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 2.8 GB | 3.4 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 1.4 GB | 1.7 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 Sep 18, 2026.
Model Card
ATQ label-gated MoE, 2.5x compressed group, v22 VLM-only conf label, SO(3)-composed merged rotation GT, RoboCasa 60k 60,000 steps, seed 42, RoboCasa 24 tasks, 2 GPUs. Labels: prehj/robocasa-conf-labels-v22 (VLM only, stride-16 anchors interpolated to every frame; frames outside the anchor span are masked out of the conf loss). confthreshold in the config is the training default -- it is an eval-only knob (the head regresses conf and never thresholds it), so override it at serve time with --conf-threshold. optimizer.pt is not included: the run finished its schedule, so there is nothing to resume; the weights are what you want.
Excerpt from the card by Jeon Hojin.
Configuration
- Architecture
- GR00T_N1_5
- Hidden size
- 2,048
- Stored precision
- bfloat16
- Model type
- gr00t_n1_5
Identity and Version
- Repository
- prehj/atq-v22so3-labelgated-2.5x-60k
- Publisher
- Jeon Hojin
- Task
- Not stated by the source
- Modality
- Other
- Library
- Not stated by the source
- Parameters
- 2.8B parameters
- Languages
- Not stated by the source
- Revision
- 2a2d874de1b712be5ed475536f804f7c4241c5f6
- First published
- 2026-09-18
- Last updated
- 2026-09-18
Files and Weights
8 files, 8.0 GB in total. The weights are 2 files totalling 8.0 GB in safetensors.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model-00001-of-00002.safetensors | Weights | 5.0 GB | 5b45ab61cc73 |
| model-00002-of-00002.safetensors | Weights | 3.0 GB | 56b892c96c6d |
| config.json | Configuration | 3.2 KB | — |
| experiment_cfg/metadata.json | Configuration | 28.7 KB | — |
| model.safetensors.index.json | Configuration | 106.3 KB | — |
| trainer_state.json | Configuration | 6.9 MB | — |
| README.md | Documentation | 871 B | — |
| .gitattributes | Repository | 1.5 KB | — |
License and Download
- License
- Not stated by the source
- Access
- Open weights, no gate
- Download size
- 8.0 GB
Released by Jeon Hojin through its official repository on Hugging Face.
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 8.0 GB |
| 16-bit | 5.7 GB |
| 8-bit | 2.8 GB |
| 4-bit | 1.4 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About atq-v22so3-labelgated-2.5x-60k
How much GPU memory does atq-v22so3-labelgated-2.5x-60k need?
About 6.8 GB at 16-bit and 1.7 GB at 4-bit: the weights (2.8B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run atq-v22so3-labelgated-2.5x-60k 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.