REDDIT · REDDIT// 37d agoBENCHMARK RESULT

Qwen 3.5 4B beats 0.8B in-browser

A LocalLLaMA post shows Qwen 3.5 0.8B and 4B running fully in-browser with WebGPU via Transformers.js, with no server inference. The author reports the 0.8B output was incorrect while 4B produced better results, and notes the 9B ONNX model was unavailable for testing.

// ANALYSIS

This is a useful real-world datapoint that tiny multimodal checkpoints can run locally in browsers, but quality still drops fast at the smallest sizes.

–WebGPU plus Transformers.js continues to make zero-backend local inference practical for demos and privacy-first apps.
–The 0.8B vs 4B gap reinforces that "runs locally" is not the same as "good enough for production tasks."
–Missing ONNX availability for 9B highlights tooling/export bottlenecks that still block fair model-size comparisons.

// TAGS

qwen-3-5transformers.jswebgpullminference

DISCOVERED

37d ago

2026-03-05

PUBLISHED

37d ago

2026-03-05

RELEVANCE

8/ 10

AUTHOR

manjunath_shiva