QNPUTER
A hybrid model that compiles a 32-bit RISC-V CPU into large model weights, adding 11 layers before Qwen's original 32 layers to achieve exact computation.
Timeline (last 90 days)
Developer DVN is Agro compiled an entire 32-bit CPU into the model weights of Qwen 3.5; the CPU was not trained, and the weights were directly computed by a compiler and filled in.
The Qwen 3.5 model with the compiled-in CPU got 39 of the same 40 multi-digit multiplication problems correct.
QNPUTER is a 32-bit RISC-V CPU placed in front of Qwen's original 32 layers, making the whole model 43 layers; the original Qwen weights are unchanged, the 11 new layers have about 1.04 million non-zero weights, while Qwen itself has about 4.1 billion parameters.
QNPUTER passes the official RISC-V test suite, and the author used 61 programs to confirm that its output is completely consistent with the reference machine and the QEMU emulator, with every token aligned with the reference machine.
The 11 layers newly added by QNPUTER, taken on their own, can run a stripped-down version of Python.
In a comparison on 40 large-number multiplication problems, the original Qwen 3.5 got 22 right, the version equipped with QNPUTER got 39 right, and the one connected to real Python got 40 right; the results returned by QNPUTER were all correct, and the errors originated in Qwen itself, such as dropping signs when copying answers or calling 44 times repeatedly without finishing.
With the QNPUTER approach, on an NVIDIA L4 GPU a single calculator call takes half a second to one second using a dedicated acceleration program; if only ordinary model code is used, it is more than ten times slower.
With QNPUTER, the model file goes from 9.3GB to 18.8GB, and most of the increase is padding zeros.
The author released an experimental version: running a small language model with 260,000 parameters as a C program on QNPUTER, with the weights loaded into CPU memory as data, takes about 41 seconds to generate each token.
Developer Drew Brionig used QNPUTER's compiler to compile the NES processor into a model with the same structure and ran NES games.
All Facts (10)
Developer DVN is Agro compiled an entire 32-bit CPU into the model weights of Qwen 3.5; the CPU was not trained, and the weights were directly computed by a compiler and filled in.
50%UnverifiedThe Qwen 3.5 model with the compiled-in CPU got 39 of the same 40 multi-digit multiplication problems correct.
50%UnverifiedQNPUTER is a 32-bit RISC-V CPU placed in front of Qwen's original 32 layers, making the whole model 43 layers; the original Qwen weights are unchanged, the 11 new layers have about 1.04 million non-zero weights, while Qwen itself has about 4.1 billion parameters.
50%UnverifiedQNPUTER passes the official RISC-V test suite, and the author used 61 programs to confirm that its output is completely consistent with the reference machine and the QEMU emulator, with every token aligned with the reference machine.
50%UnverifiedThe 11 layers newly added by QNPUTER, taken on their own, can run a stripped-down version of Python.
50%UnverifiedIn a comparison on 40 large-number multiplication problems, the original Qwen 3.5 got 22 right, the version equipped with QNPUTER got 39 right, and the one connected to real Python got 40 right; the results returned by QNPUTER were all correct, and the errors originated in Qwen itself, such as dropping signs when copying answers or calling 44 times repeatedly without finishing.
50%UnverifiedWith the QNPUTER approach, on an NVIDIA L4 GPU a single calculator call takes half a second to one second using a dedicated acceleration program; if only ordinary model code is used, it is more than ten times slower.
50%UnverifiedWith QNPUTER, the model file goes from 9.3GB to 18.8GB, and most of the increase is padding zeros.
50%UnverifiedThe author released an experimental version: running a small language model with 260,000 parameters as a C program on QNPUTER, with the weights loaded into CPU memory as data, takes about 41 seconds to generate each token.
50%UnverifiedDeveloper Drew Brionig used QNPUTER's compiler to compile the NES processor into a model with the same structure and ran NES games.
50%