A project called nesformer embeds the processor and memory of the Nintendo Entertainment System directly into the weights of a standard Qwen3.5 language model, producing a model that can run NES games without any training.
Project announcement · Neural computing and emulation
A language model whose weights are compiled to run NES games
The nesformer project embeds the Nintendo Entertainment System's processor and memory into the weights of a standard Qwen3.5 model. The weights are calculated by a compiler rather than learned from data, and the project says the model can run games without any training.
How the claimed pipeline works
STEP 1
NES CPU and memory logic is defined as a program
→
STEP 2
A compiler calculates each weight of a Qwen3.5 model
→
STEP 3
During play, the model predicts what the CPU does at each step
→
RESULT
The NES game runs on the model
The source text is cut off before describing how inputs, memory and the display are handled.
What the announcement claims
No conventional training was used
Weights are derived deterministically from a program
The NES is a small, well-documented test case
The model runs NES games directly
What remains unverified
Accuracy of the emulation
Speed and performance figures
Reproducible code and technical detail
Scaling to more complex systems
Key point
The usual relationship between training and computation is inverted: the weights come from a program, not from data.
Training
None, per the project
Weights
Calculated by a compiler from the NES program
Independent evaluation
Not yet available
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...