A compressed allenai/MolmoAct2-SO100_101, fine-tuned on an SO-101 cube pick-and-place dataset and built for on-device deployment (Qualcomm IQ-9075, 2× Hexagon v73 NSP).

Two compressions are applied on top of the base checkpoint:

	what	effect
layer prune	vision tower 25 → 10 blocks	5.44B → 5.22B parameters
token prune	grid token sampler, 16 tokens per crop	LLM prompt 496 → 136 tokens

Both are healed by training — the tower by embedding distillation, the token sampler by the task fine-tune (it ships with no pretrained weights and is useless without one).

How it was built

Stage 1 — depth prune + embedding distillation. The vision tower keeps blocks [0,1,2,3,5,6,7,8,9,24], chosen by per-layer SNR (10·log10(E‖in‖² / E‖out−in‖²); a high value means the block barely changes the signal). Student block j inherits teacher block keep[j]. The connector taps move from [-3,-9] to [-1,-4] so the deep tap stays on the same teacher layer (24) — which is why 24 is pinned in the keep set.

The pruned tower is then distilled against the teacher's connector output with relative MSE + (1 − cosine). The MSE is normalised by target energy because these embeddings carry massive activations (~1e2–1e3) and a raw MSE (~1e5) would drown the cosine term.

Held-out cosine, measured on an episode-level holdout rather than the training batch:

warm start 0.13441 → step 3000 0.95149 (12 evaluations, monotone)

Stage 2 — task fine-tune. 6000 steps, effective batch 28 (4 × 7 GPUs), LoRA r=64 on the VLM with a fully trainable action expert, use_grid_token_sampler=true so the sampler is trained from its random initialisation. LoRA is merged into the base weights in this checkpoint.

Measured results

Open-loop action-chunk prediction on a held-out episode tail: 200 frames, identical frames for every row (the hold baseline is 5.968 across all four, which is how you can tell). Frames whose ground-truth chunk barely moves — episode-end padding — are excluded, since "hold still" scores them perfectly and they say nothing about a policy.

checkpoint	MAE ↓	vs hold-still	step 0	step 29
base teacher, recalibrated, no fine-tune	10.125	1.696	4.333	14.828
10L student, distilled only	10.550	1.768	4.494	15.434
teacher fine-tuned — reference ceiling	4.737	0.794	2.813	6.200
this model	5.389	0.903	2.985	7.514

Task fine-tuning is what matters: it takes the teacher from 10.125 to 4.737 (−53%). Compression then costs +13.8% on top of that, and the compressed model still beats holding still (0.903 < 1.0).

Error grows along the chunk, so the cost depends on how much of the 30-step chunk you execute before re-inferring:

steps executed	teacher	this model	gap
5 of 30	3.26	3.51	+7.7%
10 of 30	3.68	4.01	+9.1%
30 of 30	4.74	5.39	+13.8%

At 30 fps a 30-step chunk is a 1-second budget and the deployment pipeline runs well inside it, so a shorter n_action_steps is a real lever rather than a compromise.

Limitations — read these

Open-loop MAE is a proxy, not a success rate. There is no simulator for this robot, so no closed-loop number exists. The upstream compression study on the LIBERO track explicitly found that its distillation cosine did not predict downstream success (two 0.970 configurations scored worse than a 0.956 one) — treat any single scalar here with the same caution.
The two compressions are not separated. The +13.8% is layer pruning and token pruning together against a reference that has neither. Which dominates has not been measured.
norm_stats.json was regenerated from the training dataset. The base checkpoint's joint calibration is offset from this dataset by roughly 153° on shoulder_lift and 103° on elbow_flex; feeding SO-101 states through the original statistics saturates the normaliser and makes the model return its own q01 vector regardless of the image. If you use this model on a different SO-101 setup, regenerate the statistics again.
Trained for one task family — 10 cube pick-and-place instructions, 624 episodes, 0.94 epochs. The reference recipe's own data shows fine-tuning had not plateaued at that budget (its 12-layer student went 88.5% → 94.5% between 2k and 5k steps), so this is likely under-trained.

Usage

python

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

repo = "ukcastle/MolmoAct2-SO101-10L-VTP"
processor = AutoProcessor.from_pretrained(repo, trust_remote_code=True, extra_special_tokens={})
model = AutoModelForImageTextToText.from_pretrained(
    repo, trust_remote_code=True, dtype=torch.bfloat16
).to("cuda").eval()

out = model.predict_action(
    processor=processor,
    images=[top_rgb, wrist_rgb],          # PIL or RGB arrays
    task="Pick up the large blue cube and put it in the box.",
    state=joint_positions,                 # raw 6-dim, degrees
    norm_tag="so100_so101_molmoact2",
    inference_action_mode="continuous",
    enable_depth_reasoning=False,
    num_steps=10,
)
actions = out.actions                      # [1, 30, 6], robot scale

use_grid_token_sampler is baked into config.json and must stay on. The token sampler is inlined into this checkpoint's own modeling_molmoact2.py and processing_molmoact2.py, so trust_remote_code=True is enough and lerobot does not need to be installed. A checkpoint that carries sampler weights next to base-model remote code instead loads without any error and is silently wrong: the 8 sampler tensors are dropped as UNEXPECTED and a 32-token-trained model runs at 392 tokens.

Licence

The base model allenai/MolmoAct2-SO100_101 carries no declared licence, on its model card or as a hub tag, so no licence is asserted here either. The training data is Apache-2.0. Confirm the upstream terms with AllenAI before redistributing this checkpoint or using it commercially.