Penguin-VL¶

Development line: project:penguin-vl · thread penguin-vl-public-release
Last event: 2026-03-09 · 1 dated since 2026-03-09 · Researched: 2026-09-05 · confidence: high

What it is¶

Penguin-VL is a Qwen3-based VLM family for practitioners who need local multimodal inference rather than a hosted vision API.

• Image, document, chart, multi-image, and video prompts
• Transformers, Gradio, and vLLM serving paths
• 2B and 8B checkpoints, plus a 0.4B vision encoder

Development line¶

• 2026-03-09 — Penguin-VL public project resources were shared. On 2026-03-09, Penguin-VL linked the project's website, source repository, Hugging Face collection, and interactive Hugging Face Space. This public milestone allows evaluating the project, finding its published assets, and testing its demo. The dated links alone do not establish the exact model version, technical claims, license, or release scope.

What changed¶

2026-03-09 — Tencent released Penguin-VL-2B, Penguin-VL-8B, Penguin-Encoder, inference code, a vLLM plugin, and a Gradio demo. 2026-03-17 — training code was released. 2026-03-20 — Penguin-Recap-I, the accompanying reconstructed image-training data, was released. 2026-03-26 — lmms-eval gained Penguin-VL benchmark support. 2026-03-30 — Penguin-Recap-V, video data with dense time-, paragraph-, and video-level annotations, was released.

How to use this¶

As of 2026-03-09, evaluate Penguin-VL through its website, source repository, Hugging Face collection, and demo. Check capabilities, versions, licensing, and deployment requirements in those primary resources before relying on it.

1. Create a clean environment with Python 3.11.13, PyTorch 2.5 or later, and CUDA 11.8 or later. Install the repository requirements, then run the supplied Transformers example with either tencent/Penguin-VL-2B or tencent/Penguin-VL-8B. — https://github.com/tencent-ailab/Penguin-VL
2. For direct Python integration, load tencent/Penguin-VL-8B through AutoModelForCausalLM and AutoProcessor with trust_remote_code=True. Pass image, video, or text inputs through the processor. — https://huggingface.co/tencent/Penguin-VL-8B
3. For an internal visual test UI, launch inference/launch_gradio_demo.py --model-path tencent/Penguin-VL-2B or the 8B checkpoint, then use the local interface. — https://github.com/tencent-ailab/Penguin-VL
4. For an OpenAI-compatible serving endpoint, install the separately versioned vLLM stack. Start the project plugin with python -m penguinvl.plugin.vllm serve tencent/Penguin-VL-8B. — https://github.com/tencent-ailab/Penguin-VL

Best practices¶

• Keep Transformers inference and vLLM inference in separate environments: the documented vLLM 0.11.0 path requires PyTorch 2.8, while the standard path recommends PyTorch 2.5 or later. — https://github.com/tencent-ailab/Penguin-VL
• Install vLLM before Flash Attention when using the vLLM path to avoid dependency conflicts. This is the project's documented order. — https://github.com/tencent-ailab/Penguin-VL
• Use the provided inference notebook and test representative OCR, document, chart, and video inputs before adopting benchmark claims for production routing. — https://github.com/tencent-ailab/Penguin-VL

Superseded by this¶

• 2026-03-17 — the 2026-03-09 inference-only release is incomplete guidance for practitioners who need to reproduce or adapt training; training code is now available.
• 2026-03-20 — guidance that Penguin-VL training data was unavailable is obsolete for the image side; Penguin-Recap-I is available.
• 2026-03-30 — image-only training-data guidance is incomplete for video work; Penguin-Recap-V adds multi-granularity video annotations.

Still unknown¶

• The hosted Hugging Face Space displayed a configuration error when read on 2026-09-05; local Gradio remains the documented demo route.
• We found no independent deployment or cost measurements in the consulted first-party sources.
• We skipped the original event’s hf.ru redirect because its destination and publication context were not verified.

Sources¶

Agent brief¶

• Subject: project:penguin-vl, thread penguin-vl-public-release, 1 dated events 2026-03-09 → 2026-03-09.
• Practical note: As of 2026-03-09, practitioners can begin evaluating Penguin-VL through its linked website, source repository, Hugging Face collection, and demo; verify capabilities, versions, licensing, and deployment requirements from those primary resources before relying on it.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.