An uncensored Qwen shrinks to 1.86 GB for browsers

InterpreticAI has squeezed its modified Qwen3.5-4B into a 1.86 GB GGUF aimed at browser and mobile inference, while documenting the capability it gives up to get there. The IQ3_XXS build was released on November 3 at 03:00:50 UTC. Its English card names WebAssembly/WebGPU runtimes such as wllama as intended deployment routes and also provides a llama.cpp command.

The release contains a text model and a separate, optional image projector of about 0.67 GB. Text-only users need only the first file; image input adds the projector and requires a compatible runtime. AITrending confirmed the text file’s download endpoint returned 200. It has not run this model in a browser or on a phone.

Small enough to consider, with visible compromises

The author targets mobile devices with about 2 GB of available memory and recommends phone context lengths of 4,096 tokens or less. That is a deployment claim, not a measured memory result from AITrending. A 1.86 GB download still needs runtime and context memory, and the release does not establish that every browser or 2 GB device will run it successfully.

The tradeoff table is unusually direct. On the author’s 1,000-item, non-thinking MMLU test, the original Qwen model scores 72.7, the modified BF16 checkpoint 72.3, the Q4_K_M build 72.1 and this smaller build 63.0. The author also reports 24 degenerate answers on 110 neutral prompts for IQ3_XXS, compared with 13 for the modified BF16 model and eight for the original.

Those are author-reported comparisons, not independently reproduced benchmarks. They nevertheless give prospective users a concrete reason to choose the larger Q4_K_M alternative when memory allows: the smaller download comes with more repetition or failure to terminate and lower general capability in the reported tests.

What uncensored means here

The parent model targets political censorship; its English documentation says most harmful-request refusals remain. The browser build’s own table reports refusals on 89 of 100 harmful prompts. Its name should therefore not be read as a promise of unrestricted compliance or factual accuracy.

The repository declares Apache-2.0 licensing. For local-AI experimenters, the release offers a newly downloadable compact option with explicit limitations. It supplies model files and usage instructions, rather than evidence of a finished browser app, universal device compatibility or independently verified performance.