DeepHorus is an innovative tech media and intelligence platform focused on AI, blockchain, startups, and emerging technologies.
Published on

Drew Breunig embeds NES 6502 CPU emulator in Qwen3.5-4B weights

Authors
  • avatar
    Name
    DeepHorus
    Twitter

Drew Breunig has embedded a working emulator for the Nintendo Entertainment System’s 6502 CPU into the weights of Qwen3.5-4B. The project, nesformer, lets users run NES ROMs such as Super Mario Bros inside Qwen, very slowly. Breunig announced it on 30th October 2026 via his Link Blog.

The effort was inspired by qwenputer, a related project by Divyansh Agrawal. Breunig used Opus 5.5 prompts to investigate emulating the NES chip and RAM, similar to FPGA-based MISTR projects. He also asked what model size would be needed and whether smaller, faster models could reach standard clockspeed.

In a Hacker News comment, Breunig shared the prompts that guided the build. The first told Opus 5.5 to read the qwenputer article and relevant source code before assessing feasibility. The second said: “I think we should build it! Just for fun. It’s okay if it’s not fast.”

The resulting system stores the emulator in the model’s weights rather than conventional code. Users can point a command-line interface at a ROM, according to Breunig’s stated goal. The approach echoes FPGA hardware emulation but moves the logic into neural-network parameters.

The project matters because it tests how far model weights can be pushed beyond text generation. It also shows LLMs can help design unusual, low-level systems when guided by detailed prompts. The slow speed suggests performance remains a major constraint.

Opus 5.5 apparently found and used NES CPU conformance tests released 15 years ago by Christopher Pow. Those tests likely helped validate the emulator’s behavior. Public reaction has focused on the work’s playful, experimental nature.

Breunig’s project does not promise practical NES emulation at full speed. It offers a proof of concept for embedding hardware emulation inside model weights. Future work could explore smaller models and faster execution, but no timeline has been announced.

Read original article →Source: simonwillison