Same prompt. Two models.
Ask anything — the base model and the Tileward-compressed model answer side by side. Same weights, tessellated and quantized: smaller on disk and VRAM, quality held.
Same model, two builds · Tileward = Llama-3.2-3B compressed 2.3× (smaller & faster on-device) · perplexity shown per model, lower = better · conversations saved