---
title: LoopVL
emoji: 🔁
colorFrom: purple
colorTo: yellow
sdk: gradio
sdk_version: 6.30.0
app_file: app.py
short_description: Recurrent VLM with a controllable compute budget
python_version: "3.12"
startup_duration_timeout: 30m
---

# LoopVL: Recurrent Vision-Language Model

**LoopVL** ([TierFlow/LoopVL](https://huggingface.co/TierFlow/LoopVL)) is a ~1.6B
vision-language model that reuses its 16-layer transformer stack in nested loops
instead of stacking 128 layers. The distinctive control this demo exposes is the
**recurrent inference budget**: the H (high) and L (low) cycle counts decide how
many times the stack is applied per forward pass — unrolled depth
`16 × H × (L+1)` effective layers. The schedule can be changed at runtime without
reloading weights.

Ask any question about an image, pick a schedule, and compare answers across
compute budgets.

- **Model**: [TierFlow/LoopVL](https://huggingface.co/TierFlow/LoopVL) (BF16
  vision/language cores + FP32 projector, visual gate and anchor)
- **Code**: [Tier-Flow/LoopVL](https://github.com/Tier-Flow/LoopVL) — the
  inference runtime (`hrm_penguin`, `penguin_encoder`, `modeling_loopvl.py`,
  `metadata.py`) is vendored here from that Apache-2.0 repository; the Penguin
  vision encoder keeps its Tencent `LICENSE.txt` alongside.
- **Examples**: images from
  [linoyts/repo-to-space-example-inputs](https://huggingface.co/datasets/linoyts/repo-to-space-example-inputs)
  (free to use, CC0-style).

## How it works

1. The image is resized under a dynamic-resolution policy (256–2048 visual
   tokens) and patched by the PenguinVL vision encoder.
2. Visual tokens are injected into the text sequence; the HRM-Text recurrent
   LM loops its stack `H × (L+1)` times per forward pass, re-injecting an
   anchored version of the visual tokens after each high cycle, gated by a
   learned saturation-aware gate.
3. Decoding is greedy, recomputing the full sequence per token (no KV cache).

## Reproducing the reference

The `answer` endpoint follows the authors' `runtime/infer.py` one-to-one:
same prompt template (`condition\nImage:\n` + placeholders + `\nInstruction:\n…`),
same greedy decode, same `decode(skip_special_tokens=True).strip()` postprocess.