I compiled a browser-runnable version of Qwen2.5-0.5B-Instruct using MLC-LLM/WebLLM to support the RecursiveMAS framework locally.
The model is patched to expose its last-layer hidden states, enabling client-side experiments with latent-space communication.
This setup allows developers to run the recursion loops entirely in JavaScript without relying on external API infrastructure.
Initial tests mapping to the RecursiveMAS approach show the some token reductions and speedups directly in the browser. I am in process of testing it further and would love to get some feedback
I have published the weights and implementation details on Hugging Face here: https://huggingface.co/VishalMysore/RecursiveMAS-0.5B-MLC
I compiled a browser-runnable version of Qwen2.5-0.5B-Instruct using MLC-LLM/WebLLM to support the RecursiveMAS framework locally.
The model is patched to expose its last-layer hidden states, enabling client-side experiments with latent-space communication.
This setup allows developers to run the recursion loops entirely in JavaScript without relying on external API infrastructure.
Initial tests mapping to the RecursiveMAS approach show the some token reductions and speedups directly in the browser. I am in process of testing it further and would love to get some feedback
I have published the weights and implementation details on Hugging Face here: https://huggingface.co/VishalMysore/RecursiveMAS-0.5B-MLC