开发者 Divyansh Agrawal 发布了名为 nesformer 的项目，该项目将 NES 6502 CPU 的逻辑编译进 Qwen3.5 模型的权重中 1。这一实现基于 Qwen3_5ForCausalLM 检查点，所有权重均由编译器计算得出，未进行人工训练 1。项目采用 MIT 许可证，并继承了此前 Divyansh Agrawal 开发的 Qwenputer 1。
nesformer 的核心机制是利用模型的 Gated DeltaNet 状态充当 NES 机器的内存，涵盖 RAM、寄存器和 ROM 1。通过预测 CPU 执行日志，该模拟器无需训练即可运行游戏 1。在验证测试中，所有 CPU token 均来自模型生成，且每一帧输出都与常规 6502 模拟器的结果完全一致 1。
尽管实现了精确的帧匹配，但性能方面存在显著差距 1。默认引擎在单核 CPU 上的速度约为每秒 1,650 个 tokens，整体运行速度比真实 NES 硬件慢 150 至 290 倍 1。此外，兼容性目前受到限制，仅支持 NROM（iNES mapper 0）类型的卡带，尚不支持包含银行切换芯片的游戏 1。
Divyansh Agrawal has released nesformer, a project that embeds the logic of the Nintendo Entertainment System's 6502 CPU directly into the weights of the Qwen3.5 model 1. The implementation relies on a checkpoint of Qwen3_5ForCausalLM and uses no manually trained weights; instead, the parameters are derived computationally by a compiler 1. By leveraging the model's Gated DeltaNet states to function as machine memory—including RAM, registers, and ROM—the system operates without traditional training phases 1.
Although the engine runs approximately 150 to 290 times slower than actual hardware, achieving about 1,650 tokens per second on a single-core CPU, it maintains high accuracy in emulation 1. All CPU operations are predicted via token generation from the model itself, resulting in frame outputs that match those of standard 6502 emulators exactly during tests 1. However, the current version is limited to NROM cartridges (iNES mapper 0) and does not support bank-switching chips 1. Inherited from Agrawal’s earlier Qwenputer project, nesformer is distributed under the MIT license 1.
评论
还没有评论，欢迎留下第一条。