Modelpublic

webbrain-one/webbrain-vl-2-450M-onnx

sourceHugging Faceotherupdated 1mo agoView on Hugging Face

0likes334downloads

Model Card

WebBrain VL 2 450M ONNX

This is the Transformers.js/WebGPU export of `webbrain-one/webbrain-vl-2-450M` at commit 7dd6050bbca6445ce19068b986bd62ee0d2467c3. It is derived from LiquidAI/LFM2.5-VL-450M and uses the LFM Open License v1.0.

Browser-vision benchmark

On WebBrain's unchanged 100-case production-prompt suite, this deployed Q4-decoder/FP16-vision package reached 44/100 strict passes and a 76.24% mean rubric score.

[image]

Model and tested runtime	Parameter class	Strict passes	Mean rubric
Qwen 3.5 4B, MLX	4B	55/100	83.4%
WebBrain VL 2, deployed ONNX Q4/FP16	450M	44/100	76.2%
Gemma 4 E4B IT, MLX	E4B	41/100	81.3%
Qwen 3.5 2B, MLX	2B	39/100	79.1%
Qwen 3.5 0.8B	0.8B	20/100	41.6%
LFM2.5-VL-1.6B, MLX	1.6B	12/100	44.4%

All rows used the same 100 screenshots and production six-section prompt. The benchmark images, rubrics, runner, and committed result files are public. Gemma E4B is plotted at its effective 4B class. Latency is omitted because the rows used different machines, quantizations, and inference stacks.

Production variant

Component	Precision
`embed_tokens`	FP16
`vision_encoder`	FP16
`decoder_model_merged`	symmetric Q4

The vision encoder remains FP16 to preserve small browser text and GUI state details. The package includes the exact onnx/ layout, tokenizer, processor, configuration, and chat_template.jinja expected by Transformers.js.

Evaluation

The unchanged 100-case WebBrain production-prompt release gate did not pass:

—errors: 0
—strict passes: 44/100
—mean score: 0.7623922558922558
—six-section completion: 87/100
—real WebGPU smoke: passed

See eval/ for the full PyTorch, FP16 parity, Q4 ONNX, and WebGPU smoke summaries. The WebGPU result comes from WebBrain's packaged Transformers.js/ONNX Runtime Web stack; server-side ONNX Runtime alone is not treated as proof of browser compatibility.

Provenance

—merged model: webbrain-one/webbrain-vl-2-450M@7dd6050bbca6445ce19068b986bd62ee0d2467c3
—dataset: webbrain-one/webbrain-vl-2-450M-dataset@fb191a0b73e0906e6c3b9f35f6ec0dba1337b72c
—exporter: Liquid4All/onnx-export@9a23ddd23035165f7414a5de3220a51e85780f64

The complete machine-readable provenance is in export-manifest.json and every released file is covered by SHA256SUMS.

Historical fields inside immutable export and evaluation manifests retain the repository IDs that were current when the release was produced. The current public repository names are the webbrain-vl-2-450M links used in this card.

License and attribution

Retain the included LICENSE and ATTRIBUTION.md. The derivative remains subject to the LFM Open License v1.0 redistribution and attribution terms.

Repository: `webbrain-one/webbrain-vl-2-450M-onnx`