Allen Institute for AI（Ai2）
MolmoAct
8.12B Commercial use OK 8/8 verified
Facts
- Released
- 2025-08 1 Source
- Parameters
- 8.12 B 1 Source
- Architecture
- Monolithic autoregressive 0 Source
- License
- Apache-2.0 (code and weights) 2 Source
- Commercial
- Commercial use OK 0 Source
- Weights
- huggingface.co ↗ 1 Source
- Code
- github.com ↗ 1 Source
- Paper
- arxiv.org ↗ 0 Source
- Input modalities
- rgb · depth · language 0 Source
- Action space
- End-effector pose delta + gripper 0 Source
- Cross-embodiment
- MolmoAct self-collected 10k episodes (LeRobot format) + pretraining mixture 1 Source
- Pretraining data (hours)
- Unverified 0 Source
- Target forms
- Single arm 0 Source
Architecture · Molmo family: SigLIP2 vision + Qwen2.5-7B language + VQVAE depth encoding; reasons about spatial trajectory first, then outputs actions
Sources · 2 sources
release_date,license,cross_embodiment,code_url https://github.com/allenai/MolmoAct Official page · Fetched 2026-09-04 · No snapshot stored yet README: model released 2025-08-12, code released 2025-10-24; Apache-2.0; 10k self-collected robot episodes; Google Robot / Franka
params_b,weights_url,license https://huggingface.co/api/models/allenai/MolmoAct-7B-D-0812 Model hub page · Fetched 2026-09-04 · No snapshot stored yet safetensors 8,119,417,808（F32）；license: apache-2.0；base Qwen/Qwen2.5-7B + siglip2-so400m
Activity signals
Public counters fetched daily from GitHub, Hugging Face and ModelScope. Numbers only; no judgement. · fetched 2026-10-31
GitHub stars
393
+3 / 30 d
Commits · 90 d
0
last push 174 d ago
HF downloads · 30 d
462
all-time 14,754 · 53 likes
Weights last modified
2025-10-24
373 d ago
GitHuballenai/MolmoAct · 43 forks · 3 open issues · no GitHub release
ModelScopeallenai/MolmoAct-7B-D-0812 · 181 downloads · updated 2025-10-24 · same org as Hugging Face
Hugging Faceallenai/MolmoAct-7B-D-0812
stars · last 30 d
Public benchmark scores
As printed in the source; protocols differ between papers. Full table with notes on the benchmarks page.
| Benchmark | Subset | Score | Source |
|---|---|---|---|
| LIBEROMolmoAct-7B-D | avg | 86.6 | paper · Table 2, row 'MolmoAct-7B-D', column Avg |
| LIBEROMolmoAct-7B-D | goal | 87.6 | paper · Table 2, row 'MolmoAct-7B-D', column Goal |
| LIBEROMolmoAct-7B-D | long | 77.2 | paper · Table 2, row 'MolmoAct-7B-D', column Long |
| LIBEROMolmoAct-7B-D | object | 95.4 | paper · Table 2, row 'MolmoAct-7B-D', column Object |
| LIBEROMolmoAct-7B-D | spatial | 87 | paper · Table 2, row 'MolmoAct-7B-D', column Spatial |
| SimplerEnv · Google RobotMolmoAct (fine-tuned) | variant_aggregation | 72.1 | paper · Table 1, row 'MolmoAct (fine-tuned)', Variant Aggregation Avg |
| SimplerEnv · Google RobotMolmoAct (zero-shot) | variant_aggregation | 59.3 | paper · Table 1, row 'MolmoAct (zero-shot)', Variant Aggregation Avg |
| SimplerEnv · Google RobotMolmoAct (fine-tuned) | visual_matching | 71.6 | paper · Table 1, row 'MolmoAct (fine-tuned)', Visual Matching Avg |
| SimplerEnv · Google RobotMolmoAct (zero-shot) | visual_matching | 70.5 | paper · Table 1, row 'MolmoAct (zero-shot)', Visual Matching Avg |
Deployment recipe
Environment, weights, run command and the pitfalls we hit, exactly as executed on rented GPUs (venv qwen, 30.0 GB of weights, 4 pitfalls).
Open recipe →Record history
2026-09-04
Sources fetched
release_date,license,cross_embodiment,code_url · params_b,weights_url,license
Latency & cost
cost = p50 ms ÷ 1000 × rental ÷ 3600 × 1000
| Hardware | p50 ms | p95 ms | Peak VRAM | Cost per 1k inferences ¥ | samples |
|---|---|---|---|---|---|
NVIDIA RTX 4090 ¥1.68/h · GongJi Compute (2026) | 934 | 940 | 16.25 GB | 0.436 | 1 Preliminary (n<3) Maintainer measured |
NVIDIA RTX 5090 ¥2.4/h · Runyun 2.4 / GongJi 2.5 / Paratera Cloud 2.98 (2026) | 675 | 713 | 16.25 GB | 0.450 | 1 Preliminary (n<3) Maintainer measured |
NVIDIA A100 80GB no rental price | — | — | — | — no latency data | No data |
NVIDIA Jetson AGX Thor no rental price | — | — | — | — no latency data | No data |
NVIDIA H100 80GB no rental price | — | — | — | — no latency data | No data |
NVIDIA A800 80GB no rental price | 937 | 969 | 16.25 GB | — no rental price for this hardware | 1 Preliminary (n<3) Maintainer measured |
Compatibility
Colored only where we hold evidence. An amber dot means the vendor declares this form as a target; that is a claim, not a verified compatibility.
| Unitree G1 | Unitree H1 | AgiBot Genie G1 | AgileX Piper arm | SO-101 | Franka FR3 / Panda | Fourier GR-2 | WidowX 250 S | Google Robot（Everyday Robots） | ALOHA 2 / Cobot Magic (dual-arm) | Galaxea R1 Pro | Franka Emika Panda | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| MolmoAct |
Legend Officially supportedCommunity reproducedFine-tunable in theoryUnsupportedUnknown Vendor-declared target form 3 / 12 cells with a definite status
Reproduction records
Did this model run on your robot? Report it with a link (log, video or repo). Records are reviewed by the maintainers; a verified “works” or “works after fine-tuning” upgrades the matrix cell to “community verified” with your link as evidence. Anonymous is fine; nothing but the fields below is stored.
No verified record yet for this model.
Clear direction, grounded judgement.