门电路版 BitCPM4-1BBitCPM4-1B in Logic Gates
整个语言模型被现场生成为 NAND / LATCH 门电路网表,在你的显卡上用 WebGPU 逐位执行;每个 token 约需 1.4 万亿次 NAND。
单元库是预综合的固定资产;「按权重选单元与连线」由本页在浏览器里现场生成。文字 ↔ 编号由原 tokenizer 在页面里完成(宿主 I/O,不参与计算)。
带工具定义时,工具调用由门内的 K6 摘要机生成(工具不会被执行)。首次打开需下载约 750 MB 并生成电路(几分钟);之后从本机缓存启动。The whole language model is generated on the spot as a NAND / LATCH gate netlist and executed bit by bit on your GPU with WebGPU; each token takes about 1.4 trillion NAND evaluations.
The cell library is a fixed, pre-synthesized asset; “choosing cells and wiring from the weights” is generated live in your browser by this page. Text ↔ token IDs are converted in the page by the original tokenizer (host I/O, not part of the computation).
When tool definitions are given, tool calls are generated by the in-gate K6 summarizer (tools are never executed). The first visit downloads about 750 MB and generates the circuit (a few minutes); after that it starts from the local cache.
设备与准备Device & setup:: 检查中…Checking…
工具定义(可选,JSON 数组,OpenAI function 格式)Tool definitions (optional; JSON array in OpenAI function format)