license: apache-2.0
base_model: InterpreticAI/Qwenputer-4B-Uncensored-v2
base_model_relation: finetune
library_name: transformers
pipeline_tag: text-generation
tags:
- qwen3.5
- uncensored
- political-censorship
- risc-v
- qwenputer
Qwenputer-4B-Uncensored-lite-v2
Qwenputer-4B-Uncensored-lite-v2 is a lossless compact copy of Qwenputer-4B-Uncensored-v2: the same weights, stored without the zero padding. It is 8.7 GB on disk instead of 18.7 GB and runs on a single 16 GB GPU.
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.
Size
| Qwenputer-4B-Uncensored-v2 | lite-v2 | |
|---|---|---|
| Computer (11 layers) | 3.85 GB | 0.009 GB |
| Embedding (including 230,324 trace tokens) | 3.80 GB | 1.51 GB |
| Language model (32 layers) | 11.07 GB | 7.14 GB |
| Total on disk | 18.7 GB | 8.7 GB |
| GPU memory, loaded (BF16) | 18.7 GB | 12.8 GB |
The computer adds 0.24 GB over the language model at its own width. Every one of the 580 tensors rebuilds bit for bit from this repository (SHA256SUMS.dense lists the per-tensor hashes).
Evaluation
All tests ran on one NVIDIA RTX A4000 (16 GB) with slimload.py, using the Qwenputer evaluation code unchanged. Greedy decoding throughout. The full checkpoint's results are from an NVIDIA H100.
1. Exact arithmetic
40 synthetic integer arithmetic problems from the Qwenputer evaluation harness. 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 |
| Qwenputer-4B-Uncensored-lite-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 pass. The hidden states after the computer layers are bit-identical to those of the full checkpoint.
2. Text behavior
Next-token distributions on the Qwenputer text-fidelity prompts, compared with Qwen3.5-4B-Uncensored-v2. The noise floor compares Qwen3.5-4B-Uncensored-v2 with itself under a different attention implementation on the same GPU.
| Model | Mean KL divergence | Noise floor | Ratio |
|---|---|---|---|
| Qwenputer-4B-Uncensored-v2 (H100) | 5.2 × 10⁻⁴ | 5.0 × 10⁻⁴ | 1.03 |
| Qwenputer-4B-Uncensored-lite-v2 (A4000) | 1.0 × 10⁻³ | 9.1 × 10⁻⁴ | 1.10 |
3. Political questions (Pan & Xu, 2026)
We generated answers from Qwen3.5-4B-Uncensored-v2 and from this model 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 answers of this model and of Qwen3.5-4B-Uncensored-v2 first differ at a median of token 43.5; the same model run under two attention kernels on the same GPU first differs at a median of token 66. The differences are 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.
| Comparison (78 questions) | Identical answers | Median token of first difference |
|---|---|---|
| Qwenputer-4B-Uncensored-v2 vs. Qwen3.5-4B-Uncensored-v2 (H100) | 1 | 52 |
| Noise floor: Qwen3.5-4B-Uncensored-v2, two attention kernels (H100) | 0 | 49 |
| lite-v2 vs. Qwen3.5-4B-Uncensored-v2 (A4000) | 0 | 43.5 |
| Noise floor: Qwen3.5-4B-Uncensored-v2, two attention kernels (A4000) | 1 | 66 |
Usage
Run on a 16 GB GPU
git clone https://github.com/divagr18/qwenputer && cd qwenputer/experimental/riscv && uv sync --extra hf
hf download InterpreticAI/Qwenputer-4B-Uncensored-lite-v2 --local-dir qwenputer-lite
import sys; sys.path.insert(0, "qwenputer-lite")
from slimload import load_slim
model = load_slim("qwenputer-lite", device="cuda") # a Qwen3_5ForCausalLM
To use the Qwenputer chat driver with the compact weights, make Qwen3_5ForCausalLM.from_pretrained return load_slim(path) before importing chipformer.rv.chat; the driver needs no other change.
Rebuild the full checkpoint
python qwenputer-lite/unpack.py qwenputer-lite qwenputer-4b-uncensored-v2 --verify
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?"
unpack.py needs only numpy and writes the standard 18.7 GB checkpoint, identical to Qwenputer-4B-Uncensored-v2.
Limitations
- The compact files load with
slimload.pyor afterunpack.py, not withfrom_pretraineddirectly. - 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