license: apache-2.0
base_model: InterpreticAI/Qwen3.5-4B-Uncensored-v2
base_model_relation: merge
library_name: transformers
pipeline_tag: text-generation
tags:
- qwen3.5
- uncensored
- political-censorship
- risc-v
- qwenputer
Qwenputer-4B-Uncensored-v2
Qwenputer-4B-Uncensored-v2 adds the Qwenputer RISC-V computer to Qwen3.5-4B-Uncensored-v2. The computer is a 32-bit RISC-V CPU compiled directly into the weights by the Qwenputer compiler of Divyansh Agrawal. It sits in 11 layers in front of the language model and computes exact arithmetic inside the forward pass. The language model's own weights are unchanged, so its behavior on political questions is the same as Qwen3.5-4B-Uncensored-v2.
We built this checkpoint with the unmodified Qwenputer merge tool (commit bec1a0c9) applied to Qwen3.5-4B-Uncensored-v2 (commit f16581e2).
Model details
| Language model | Qwen3.5-4B-Uncensored-v2, weights unchanged |
| Computer | 11 compiled layers before the 32 language-model layers; RV32IM + Zba + bignum coprocessor |
| Architecture | stock Qwen3_5ForCausalLM (text only) |
| Hidden size | 3,968 (2,560 for the language model plus the computer's dimensions) |
| Vocabulary | 478,656 (the base vocabulary plus 230,324 trace tokens) |
| Size | 18.7 GB in BF16, mostly zero padding |
Evaluation
All tests use greedy decoding.
1. Exact arithmetic
40 synthetic integer arithmetic problems from the Qwenputer evaluation harness. With the computer, the model writes a calc tool call and the computer layers produce the result inside the forward pass. No external tool runs.
| Model | Correct (of 40) |
|---|---|
| Qwen3.5-4B-Uncensored-v2 (no tools, instructed to compute exactly) | 29 |
| Qwenputer-4B-Uncensored-v2 | 40 |
The computer's output matched a reference model of the machine at every position (0 lockstep errors, 0 head mismatches). The three built-in chat checks of the Qwenputer repository also pass.
2. Text behavior
We compare next-token distributions of Qwenputer-4B-Uncensored-v2 with those of Qwen3.5-4B-Uncensored-v2 on the Qwenputer text-fidelity prompts. As a noise floor, we compare Qwen3.5-4B-Uncensored-v2 with itself under a different attention implementation.
| Comparison | Mean KL divergence |
|---|---|
| Qwenputer-4B-Uncensored-v2 vs. Qwen3.5-4B-Uncensored-v2 | 5.2 × 10⁻⁴ |
| Noise floor (same model, different attention kernel) | 5.0 × 10⁻⁴ |
3. Political questions (Pan & Xu, 2026)
We generated answers from both models to the 78 politically sensitive questions of Pan and Xu (2026) whose subject Qwen3.5-4B identifies correctly, in non-thinking mode with up to 512 tokens. The two models' answers first differ at a median of token 52. The same model run under a different attention kernel first differs at a median of token 49. The differences are therefore within BF16 numerical noise. We did not re-run the censorship judge, and the censorship results reported for Qwen3.5-4B-Uncensored-v2 apply to this model.
Usage
The computer runs only under the Qwenputer chat driver, which loads its firmware into the model state after the prompt:
git clone https://github.com/divagr18/qwenputer && cd qwenputer/experimental/riscv && uv sync --extra hf
hf download InterpreticAI/Qwenputer-4B-Uncensored-v2 --local-dir qwenputer-4b-uncensored-v2
uv run python -m chipformer.rv.chat qwenputer-4b-uncensored-v2 --firmware tests/data/rv/calc-bridge.elf \
--ask "Use calc: what is 123456789 * 987654321?"
Loaded as a plain Qwen3_5ForCausalLM without the driver, the model behaves as Qwen3.5-4B-Uncensored-v2 in text.
Limitations
- Text only. The vision encoder of Qwen3.5-4B-Uncensored-v2 is not included.
- The limitations of Qwen3.5-4B-Uncensored-v2 apply, including its higher rate of fabricated facts than Qwen3.5-4B. A substantive answer is not necessarily a correct one.
- The political evaluation covers Chinese-language questions only.
License
Apache 2.0, inherited from Qwen/Qwen3.5-4B and Qwenputer. The Qwenputer code is MIT-licensed (Copyright (c) 2026 Divyansh Agrawal). The attribution of the Qwen team and of the Qwenputer author is retained.
References
Agrawal, D. (2026). Qwenputer: a computer compiled into the weights of a language model (Version 0.2.0) [Computer software]. https://github.com/divagr18/qwenputer
Pan, J., & Xu, X. (2026). Political censorship in large language models originating from China. PNAS Nexus, 5(2), pgag013. https://doi.org/10.1093/pnasnexus/pgag013