| title: NCM chat (experimental, rough) | |
| emoji: 💬 | |
| colorFrom: gray | |
| colorTo: blue | |
| sdk: static | |
| pinned: false | |
| license: mit | |
| # NCM chat (experimental, rough) | |
| A small chatbot built on **Next-Concept Modeling (NCM)**, Stanford Hazy Research (arXiv 2610.14162, code github.com/HazyResearch/ncm). | |
| - Base: the released XSum NCM checkpoint (frozen multiscale VQ tokenizer plus prefix-conditioned generator). No chat checkpoint is released, so only the generator was fine-tuned for a short time on short public dialogue and instruction data (everyday-conversations, dolly, alpaca). | |
| - Runs fully in your browser with onnxruntime-web. No server and no data leaves your device. | |
| - Weights are stored fp16 and cast to fp32 at load (about 375 MB download, roughly 1 GB RAM). Desktop browser recommended. Phones are untested. | |
| - Export check: on 8 test prompts, greedy replies from the ONNX model are identical to the PyTorch model, and teacher-forced logits match (max diff 0.0026, 0/1600 argmax mismatches). | |
| - It is a rough research demo. Replies are short and often wrong or repetitive. It is not a capable assistant. | |