What kind of black magic is squeezing a 32-bit RISC-V CPU into 4B model weights? In this issue, we introduce Qwen3.5-4B-Qwenputer, which allows LLMs to perform precise large-number arithmetic and even execute C code without a sandbox. We also cover maya2-native, a voice cloning model that runs locally with low latency. Read on to find out how it outperforms commercial voice APIs!
Featured Projects
1. divagr1925/Qwen3.5-4B-Qwenputer: A Geeky Model Running a RISC-V CPU Inside Its Weights
- What is it: This is a unique 4B parameter model licensed under apache-2.0 and provided in safetensors format. The craziest part is that the author compiled a complete 32-bit RISC-V CPU and embedded it directly into the weights of Qwen3.5-4B, using the model's own Gated DeltaNet state as RAM to run.
- Use Cases:
- Scenario 1 (Sandbox-free precise math): When you need precise calculations (such as multiplying or dividing large numbers like 123456789 * 987654321) but do not want to rely on an external Python execution environment or set up a secure sandbox, this model can perform precise calculations directly inside the LLM, avoiding the calculation errors common in traditional LLMs.
- Scenario 2 (Deterministic logic control): If you need completely deterministic, error-free logic control during the LLM's forward pass, this model can directly execute compiled C or MicroPython code to achieve this.
- Scenario 3 (Persistent file system across conversations): During multi-turn conversations, you can use the Gated DeltaNet state as a persistent file system to save file states between turns and load them directly in the next turn, without needing to maintain an external database.
- Can it run?: According to the official README, the model file size is 18.8 GB in BF16 precision, though most of it is zero-padded. Its CPU runs on a Triton executor. The official README notes that when running on an NVIDIA L4 GPU, it can process about 3,200 trace tokens per second.
- Popularity: It currently has a trending score of 89 on Hugging Face, with 572 downloads and 89 likes.
- Comparison: The author notes that compared to the tool-free Qwen3.5-4B base model, Qwenputer's score on the DeepMind Mathematics benchmark jumped from 55%-90% to an impressive 97.5%-100%. In the NUPA test, it achieved a perfect 100% score, compared to just 48.6% for the base model. The author states that its performance is equivalent to calling a real Python interpreter on the host side, but it runs entirely within the weights.
- How to get it: The official README explains that you can clone the project via
git clone https://github.com/divagr18/qwenputerand useuv run python -m chipformer.rv.chatto load the firmware and model to start interacting with it. - Link: divagr1925/Qwen3.5-4B-Qwenputer
2. maya-research/maya2-native: A Local, Low-Latency Speech Synthesis and Voice Cloning Powerhouse
- What is it: This is an open-source speech synthesis and voice cloning model released by maya-research. It features a developer-friendly apache-2.0 license and safetensors format, focusing on excellent local execution efficiency.
- Use Cases:
- Scenario 1 (Local real-time voice dialogue systems): If you want to set up a low-latency voice agent or Interactive Voice Response (IVR) system locally, this model can perform real-time streaming speech synthesis with 80ms audio frames, providing extremely smooth voice interactions without noticeable lag.
- Scenario 2 (Zero-shot cross-lingual voice cloning): Content creators who want to make their voice speak other languages only need to provide a clean 3 to 10-second audio clip of a single speaker. They can then perform zero-shot voice cloning locally and use the cloned voice to read text in other languages.
- Can it run?: The official model card indicates that under a single connection (max_seats=1), the peak memory load is around 9.2 GB (this value excludes the CUDA context). The official recommendation is to use a GPU with 12 GB or more VRAM.
- Popularity: It currently has a trending score of 119 on Hugging Face, with 1.5k downloads and 120 likes.
- Comparison: The official model card cites VoiceArena's Hindi real-time TTS leaderboard data as of October 27, 2026, where Maya 2 Native ranked 2nd, outperforming well-known commercial and peer models such as Cartesia Sonic-3.5, Smallest AI Lightning 3.1 Pro, and xAI Grok TTS.
- How to get it: The official README explains that you can install it via
pip install "maya3[server] @ git+https://github.com/MayaResearch/maya2-native"and runpython -m maya3.serverto start the local HTTP and WebSocket servers. - Link: maya-research/maya2-native
Conclusion
The two local models introduced in this issue perfectly showcase the geek spirit of the open-source community pushing performance to its absolute limits. Whether it is squeezing a CPU directly into LLM weights to run deterministic logic, or achieving ultra-low-latency voice cloning locally with just a 12 GB graphics card, both demonstrate the infinite potential of local applications. If you are itching to try out some mind-blowing tech, we highly recommend starting with the Qwenputer project. See you in the next issue!
Comments