An independent developer has built an experimental version of Qwen3.5 4B that contains a functioning 32-bit RISC-V processor in additional model layers. The project, called Qwenputer, was described on October 29 as a demonstration that a language model’s own recurrent state can be converted into exact, addressable memory and used to execute conventional programs during inference.
The processor is not a plugin, sandbox or separate runtime. According to the project’s technical account, a compiler directly sets the added weights so that attention heads perform lookups, feed-forward components handle arithmetic, and Gated DeltaNet heads hold the processor’s memory. Qwen’s original weights remain unchanged, while the computer occupies 11 layers placed before the model’s existing 32 layers.
Qwen3.5 uses Gated DeltaNet, a linear recurrent architecture, in most of its attention layers. Each head maintains a fixed-size state matrix. Qwenputer configures the gate, write strength and keys so that individual rows can be overwritten and read exactly, turning that state into RAM. This differs from ordinary attention memory, whose key-value cache grows with the conversation.
The resulting core implements the RV32IM instruction set, the Zba extension and a small unit for large-number arithmetic. The developer reports that it passes the official RISC-V tests, agrees with QEMU on program behavior and completed 18 million instructions without an error in a free-running test. Those performance claims come from the project author and have not been independently reproduced in the supplied evidence.
Programs execute while the model generates a compact trace. Tokens identify a memory address and value, followed by a step marker that advances the processor. A small driver loads firmware into the recurrent state and keeps those trace tokens away from Qwen’s language layers. The system can run compiled C and MicroPython, although the author says it remains slow and is not a Linux-class machine.
In one reported evaluation, the embedded processor raised Qwen’s score on a long-integer arithmetic task from 48.6% to 100%, matching the result achieved when the model was given a Python tool. The developer also replayed roughly 230,000 output positions against a reference machine and reported no mismatches.
The experiment includes an in-model file store as well. It hashes names into 1,024 directory slots and distributes text across recurrent-state rows, with a stated capacity of 512 files or about 53,000 tokens. Saving and restoring the model state can preserve those files between conversations.
Qwenputer is available through model repositories alongside its code. Its significance is experimental: it shows one way deterministic computation can be compiled into a neural model rather than reached through an external tool call. Broader conclusions will depend on independent testing of the implementation, its resource costs and its behavior across hardware and workloads.
