Launches
ThinCtx says its compact Qwen3.5-4B fits 128k context in 4 GB of GPU memory
Community reports 3 sources
T
ThinCtx-Qwen3.5-4B-small
Model Oct 14, 2026
A compact Qwen3.5-4B conversion whose author reports lower GPU context memory using the dedicated ThinCtx runtime and system RAM.
Visit primary link
3 sources attached
What happened lately
ThinCtx-Qwen3.5-4B-small timeline
ThinCtx-Qwen3.5-4B-small is thinletter’s compact conversion of Qwen3.5-4B, with weights stored at about 5.4 bits per weight and a compressed context store. Its author reports 4.0 GB of GPU memory for weights and context at 128k tokens, with the full context also occupying system RAM. These memory figures have not been independently reproduced by AITrending.
The conversion requires thinctx 0.3.1 or later on Linux x86_64 with CPython 3.12, an NVIDIA GPU and PyTorch 2.11. The documentation warns that long-prompt reading is slower than the base model. Its PolyForm license includes free qualifying small-business use and additional personal, non-commercial research and evaluation permissions; larger organizations need a commercial license for production.