CAS

BitVLA

Chinese Academy of Sciences, China · June 2025

Institute of Computing Technology, Chinese Academy of Sciences

Compact VLAOpen weights, MITAlso written BitVLA 1-bit, bitvla-bitsiglipL-224px
Parameters
3 B
3.0B parameters, reported both in the paper and in the GitHub README.
GPU memory
1.4 GB
reported by the authors
Inference latency
73 ms
NVIDIA A100, end to end latency for an action chunk of K=25, corresponding to a reported throughput of 341.1 Hz.
Weights
MIT
downloadable checkpoint

What it is

BitVLA is a natively 1-bit vision language action model in which every LLM weight is ternary, restricted to minus one, zero or one. It builds on BitNet b1.58 2B4T and adds a SigLIP-L vision encoder that is itself compressed to 1.58-bit weights by a Quantize-then-Distill procedure, with a full precision teacher guiding representation alignment. At 3.0B parameters it occupies 1.4 GB of model memory against 15.4 GB for OpenVLA-OFT, and runs at 73 ms end to end on an A100 against 321 ms, while scoring 96.0 percent on the LIBERO average against OpenVLA-OFT's 97.1. The evidence base is narrow: LIBERO is the only simulation benchmark reported, real world testing is three tasks on one Franka arm, and every efficiency number comes from a datacentre GPU rather than the edge hardware the design targets. The v2 revision and the pretrained checkpoint were released in March 2026.

Architecture

Backbone
BitNet b1.58 2B4T, a natively 1-bit LLM with ternary weights, paired with a SigLIP-L vision encoder at 224x224 producing 256 visual tokens. The vision encoder is compressed to 1.58-bit weights with INT8 activations through a Quantize-then-Distill scheme in which a full precision teacher guides representation alignment.
Action head
continuous action decoding through a full precision MLP head
Action chunk
8 steps per forward pass
Parameters
3.0B parameters, reported both in the paper and in the GitHub README. Every LLM weight is ternary, restricted to {-1, 0, 1}. The action head remains full precision. The compressed vision encoder accounts for about 0.1 GB of the 1.4 GB total memory footprint. Released Hub checkpoints are stored in bf16 containers despite the ternary weight design.
Pretraining data
Vision language pretraining plus VLA pretraining on 16x NVIDIA H800 80GB GPUs for approximately two weeks. The pretrained BitVLA checkpoint was released in March 2026 alongside the v2 paper revision, which updated the reported evaluation results.
Embodiments
LIBERO (simulation), Franka Emika 7-DoF with RealSense D435i (real)

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Computed from 3 B parameters. Weights only, so treat it as a floor: activations, image encoder intermediates and the CUDA context come on top, in practice 30 to 50 percent more for inference.
PrecisionWeightsNote
fp3211 GBTraining master weights, rarely used for inference
bf16 or fp165.6 GBThe usual way these checkpoints are served
int82.8 GBQuantised, expect some loss of precision on fine motions
int41.4 GBAggressive quantisation, verify success rate before trusting it

Smallest real card that fits the bf16 weights plus a 40 percent runtime allowance: 8 GB. That is an estimate for inference, not for fine tuning. Full fine tuning also holds optimiser state and gradients, which typically costs several times the weights unless you use LoRA or a comparable adapter.

Published figures. These come from the model authors, not from this site.
Inference memory, reported1.4 GB
Inference latency, reported73 ms on NVIDIA A100, end to end latency for an action chunk of K=25, corresponding to a reported throughput of 341.1 Hz. OpenVLA-OFT measured 321 ms on the same setup, which is the basis for the 4.4x speedup claim. Note that this is a datacentre GPU measurement, not an edge device measurement. source

Model memory is 1.4 GB, which is the entire point of the design and puts the model within reach of small accelerators. However all published speed and memory numbers were taken on an A100, and no measurement on actual edge hardware is published. Pretraining used 16x NVIDIA H800 80GB for roughly two weeks.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • LIBERO object
    99%
    success rateAuthors' own evaluation. Best of the four suites and above OpenVLA-OFT at 98.4.source
  • LIBERO spatial
    96.6%
    success rateAuthors' own evaluation. OpenVLA-OFT scored 97.6.source
  • LIBERO average over spatial, object, goal and long
    96%
    success rateAuthors' own evaluation, v2 revision. Full precision OpenVLA-OFT (7.7B) scored 97.1 and pi0 (3.5B) scored 94.2 in the same table.source
  • LIBERO long
    92.8%
    success rateAuthors' own evaluation. OpenVLA-OFT scored 94.5, pi0 scored 85.2.source

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

7-DoF Franka Emika arm with a RealSense D435i camera, three tasks (grasp watermelon, flip bell, place bread) plus out of distribution variants, trained on 50 teleoperated demonstrations per task with an action chunk of 10. The paper reports BitVLA matching OpenVLA-OFT across these tasks and their out of distribution variants. Per task success rates were not recoverable from the sources checked, so treat the real world claim as a parity statement rather than a measured margin.

Fine tuning it yourself

The repository publishes separate LIBERO fine-tuned checkpoints per suite (spatial, object, goal and long), so the intended pattern is a per suite or per task fine-tune from the pretrained BitVLA checkpoint. Simulation fine-tuning used an action chunk of K=8 and real robot experiments K=10 with 50 teleoperated demonstrations per task.

Where it helps, where it does not

Strengths

  • 1.4 GB model memory, 11.0x smaller than OpenVLA-OFT at 15.4 GB, which is the largest memory reduction of any model in this comparison.
  • 73 ms end to end latency on an A100 against 321 ms for OpenVLA-OFT, a 4.4x speedup, with a reported throughput of 341.1 Hz at chunk size 25.
  • LIBERO average of 96.0 percent, within 1.1 points of the full precision 7.7B OpenVLA-OFT baseline.
  • Efficiency comes from native 1-bit training rather than post-hoc compression, so there is no separate quantisation step to tune.
  • MIT licensed with pretrained and per suite fine-tuned checkpoints published.
  • Actively maintained: the pretrained model and updated results were released in March 2026.

Limits

  • LIBERO average of 96.0 percent still trails full precision OpenVLA-OFT at 97.1, so 1-bit is near parity rather than free.
  • Real world evaluation covers three tasks on one Franka arm and is reported as parity with OpenVLA-OFT without published per task success rates.
  • All speed and memory numbers were measured on an A100, not on the edge hardware the design targets, so the edge case is argued rather than demonstrated.
  • Ternary weights only pay off with a kernel that exploits them. The released Hub checkpoints are bf16 containers, so a naive PyTorch load will not deliver the advertised footprint.
  • No SimplerEnv, Open X-Embodiment or cross embodiment generalisation results are reported, so its generalist ability is untested.
  • Evaluation is LIBERO only in simulation, which is a narrower evidence base than SpatialVLA or NORA offer.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.