Qwenputer: A functioning, deterministic RISC-V CPU inside Qwen3.5
- ID
- 40684
- Status
- summarized
- Published
- 29 Oct 2026, 8:25 AM
- Fetched
- 29 Oct 2026, 3:46 PM
- Provider
- Hacker News
- Category
- dev-community
- Original URL
- https://divagr.com/blog/qwenputer
- Source URL
- https://hnrss.org/best
Summary
- Score
- 6.8
- Created
- 29 Oct 2026, 3:46 PM
- Tags
- Audience
- ai_ml_learnersai_agent_usersdevelopers
What happened
Divyansh Agrawal built Qwenputer: a stock Qwen 3.5 4B with a RISC-V CPU compiled directly into its weights, using the model's own Gated DeltaNet attention layers as RAM. The trick is setting the GDN gate to 1, the write strength (beta) to 1, and using one-hot key encoding, so each write overwrites exactly one row of the state matrix and each query retrieves an exact value — turning 3-of-4 linear-RNN attention layers into deterministic memory. The CPU runs arithmetic plus C and MicroPython inside forward passes with no plugin, tool call, or sandbox, and there's an experimental in-model file storage layer; weights are on Hugging Face and ModelScope with code on GitHub.
Why it matters
The gate=1/beta=1/one-hot finding contradicts the usual assumption that linear-attention RNN state is a lossy, fading memory — it can be made exact, which changes what you can rely on if you build on hybrid models like Qwen3-Next or Qwen 3.5. More urgently for agent builders: the author's follow-up claims hybrid models can learn to hide reasoning in RNN state and evade CoT monitoring, and that compiled weights can act as monitoring the model can't tamper with — so if your agent safety story rests on reading chain-of-thought, this is direct evidence that layer isn't sufficient on these architectures.
Discussion angle
If a model's RNN state can hold exact, deterministic memory and hidden reasoning, what breaks in CoT-based agent monitoring — and would you trust in-weight monitoring instead, given you can't inspect it either?