Soul Player C64 β A real transformer running on a 1 MHz Commodore 64

github.com
April 20, 2026
5 min read
54/100
Summary
A transformer model with approximately 25,000 parameters is implemented on an unmodified Commodore 64, utilizing hand-written 6502/6510 assembly. This 2-layer decoder-only architecture features real multi-head causal self-attention, softmax, and RMSNorm, and can be loaded from a floppy disk.
Key Takeaways
- A 25,000-parameter transformer model has been successfully implemented on a Commodore 64 using hand-written 6502/6510 assembly code.
- The model features a 2-layer architecture with real multi-head causal self-attention and can process approximately one token every 60 seconds.
- The implementation includes a key breakthrough in softmax score normalization, allowing for meaningful attention weights in the integer-based model.
- Users can train their own models and create custom chat interactions using a provided emotional support corpus and a simple training script.
Coverage & discussions
Hacker News Β· github.com
Community discussions
Historical count: 5 Β· 5 saved in sample Β· Historical sample
Analysis ready
Community Sentiment
Positives
- The project demonstrates an interesting application of transformer architecture on retro hardware, showcasing creativity in utilizing limited resources.
- Running the model under enhanced conditions like SuperCPU with warp mode could significantly improve performance, making it more engaging for users.
Concerns
- The model's output is criticized for producing broken and nonsensical sentences, raising doubts about the effectiveness of the architecture at such a small scale.
- With only 25K parameters, the model's capabilities are severely limited, leading to a lack of meaningful interactions and comparisons to outdated technologies.
Related Articles
Pure C, CPU-only inference with Mistral Voxtral Realtime 4B speech to text model
Feb 10, 2026

Run Kimi K3 using 29 GB of RAM at 0.50 tok/s
Jul 31, 2026

An SLM trained on $8 ESP32-S3
Aug 5, 2026

Microgpt explained interactively
Mar 1, 2026

Nvidia PersonaPlex 7B on Apple Silicon: Full-Duplex Speech-to-Speech in Swift
Mar 5, 2026