
What happened
Divyansh Agrawal released Qwenputer, a stock Qwen 3.5 4B with a 32-bit RISC-V CPU compiled into its weights across 11 extra layers, using Gated DeltaNet state as RAM. It passes the official riscv-tests and has run 18 million instructions free-running without a single error.
Why it matters
That points to models that can execute exact computations internally rather than relying on outside tools, though Agrawal presents it as an experiment.
Summaries like this, every morning.
Qwenputer builds on a property of Qwen 3.5's architecture. Three out of four attention layers in the model use Gated DeltaNet, a linear RNN, and each head keeps a state matrix. Agrawal found that by setting the gate and write strength to 1 and using one-hot key encoding, those state matrices can act as exact, addressable RAM, at lower cost than using normal attention because a GDN head's state stays the same size however long the conversation gets.
With that RAM in place, no training was needed: a compiler sets every weight directly, turning CPU steps into attention heads that look things up, feed-forward units that do the arithmetic, and DeltaNet heads that hold the memory. The CPU then runs as the model decodes, with each instruction becoming a few generated tokens that form the execution trace.
Agrawal also extended the idea to storage. A filesystem kept in the DeltaNet state stores up to 512 files, about 53,000 tokens of text, and persists when state is saved and loaded into a new conversation. He frames exact control over the RNN layers as a testbed for safety research, including whether hybrid models can hide reasoning in their RNN state and whether compiled weights can serve as monitors the model cannot switch off.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
OpenAI said Apollo Research, METR, Transluce and Goodfire will work on-site at its offices, with access to unr…

Walt Disney Imagineering R&D revealed three Audio-Animatronic prototypes—Vilo Vele, Kado Bre'a, and Jibu Bre'a…

Google rolled out temporary threads in AI Mode, letting users click an icon in the upper right corner to chat…

Reson8 released Rumble-2-Flash, a 500 million-parameter transcription model supporting over 40 languages, as a…
At Microsoft Ignite 2026, Microsoft said Microsoft Perception is now generally available, combining red, blue…

At Microsoft Ignite 2026, Microsoft released a public preview of Windows Server 2028, which adds AI inference…
