WALL-OSS-FAST
X Square Robot, China · September 2025
- Parameters
- not published
- No parameter count is published for this specific checkpoint.
- GPU memory
- not published
- weights at bf16, computed
- Inference latency
- not published
- per action step
- Weights
- Apache-2.0
- downloadable checkpoint
What it is
WALL-OSS-FAST is the discrete-action sibling of wall-oss-flow, released from the same X Square Robot repository under Apache-2.0. It uses FAST tokenization to emit discrete action tokens, corresponding to what the WALL-OSS paper calls the Inspiration stage, whereas the flow checkpoint corresponds to the later Integration stage with continuous flow matching. It is useful if you want a discrete-token action head on the same Qwen2.5-VL-3B backbone and the same LeRobot-based training stack. Be aware that no parameter count, benchmark, VRAM or latency figure is published for this checkpoint specifically, so it carries markedly less evidence than the flow variant.
Architecture
- Backbone
- Qwen2.5-VL-3B
- Action head
- Discrete action tokens produced by FAST tokenization, corresponding to the paper's Inspiration stage, as opposed to the flow-matching Integration stage used by wall-oss-flow
- Parameters
- No parameter count is published for this specific checkpoint. It shares the WALL-OSS architecture and Qwen2.5-VL-3B backbone, and the sibling flow checkpoint is listed at 4B, but no source states the number for the FAST variant so it is left empty.
- Pretraining data
- Same corpus as WALL-OSS: tens of thousands of hours combining self-collected robot data across several platform classes, more than 24 open-source action datasets and multimodal VQA.
- Embodiments
- same platform classes as WALL-OSS: desktop arms, mobile stands, wheeled bi-arm systems, wheeled humanoids
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.
Nothing published.
The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.
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
No results in this category are published for this model.
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
No separate real-robot evaluation is published for the FAST checkpoint. The paper's real-robot tables report the co-trained WALL-OSS system rather than this checkpoint by name.
Fine tuning it yourself
Same wall-x training stack, LeRobot data format and FSDP training path as the flow variant. Apache-2.0, so commercial use is permitted.
Where it helps, where it does not
Strengths
- Apache-2.0 weights
- Discrete-token action head, which makes it the cheaper variant to compare against autoregressive VLAs and easier to inspect than a continuous flow head
- Released from the same repository and stack as wall-oss-flow, so switching between discrete and continuous action heads requires no new tooling
Limits
- No published parameter count, benchmark table, VRAM figure, latency or control frequency specific to this checkpoint
- The paper treats FAST tokenization as the intermediate Inspiration stage rather than the final system, so this checkpoint is best read as an ablation artefact that was released, not as the recommended deployment target
- Superseded by Wall-OSS-0.5
Sources
Everything on this page was taken from these documents. Where they disagree with what you read here, they win.
- https://huggingface.co/x-square-robot/wall-oss-fast
- https://github.com/X-Square-Robot/wall-x
- https://arxiv.org/html/2509.11766v1
- https://arxiv.org/abs/2509.11766
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.