Qwen3.5 Small tests Windows 8GB limits

Windows is not the blocker; memory is. If you want this to work, the real levers are checkpoint size, context length, and how aggressively you quantize the cache.

• –Qwen's docs recommend local runners like llama.cpp, Ollama, LM Studio, and KTransformers: https://qwenlm.github.io/blog/qwen3/
• –The Qwen3.5-4B card warns that the default 262,144-token context can OOM, so context trimming matters: https://huggingface.co/Qwen/Qwen3.5-4B
• –TurboQuant is KV-cache quantization, which helps inference memory but doesn't replace a smaller checkpoint: https://arxiv.org/abs/2504.19874
• –By inference, an 8GB Nvidia card points you toward the Small family rather than the 397B flagship.