Small dense model from Alibaba's Qwen family, suited to on-device and low-resource deployment.
The smolmaxxing project demonstrates running language models from 260K to 16B parameters on a Xiaomi TV Box S (Amlogic S7D, ~1GB usable RAM, no root) using custom kernels, smarter memory use and weight streaming. Qwen3.5-0.8B reaches 16.8 tok/s (2.2x stock), while an 8B MoE model hits 7.2 tok/s via on-demand expert paging from flash. All claims are measured back-to-back on real hardware, offering a reproducible engineering playbook for edge-side local inference.
Blogger Nathan Barry analyzes MIT CSAIL's "weight-space programming" and finds that GPT-6 Astra's weight edit on Qwen3.5-0.8B was effectively solving a linear program by trial and error, taking about 8 hours and 28 forward passes. Two roughly one-minute LP solves achieve the same result with 8x fewer changed weights, suggesting the core mechanism can be reproduced efficiently by classical optimization and offering a direct reference for weight editing and gradient-free fine-tuning.