ReAligned-Qwen3.5 is a family of Qwen3.5-based language models realigned to reduce China-state ideological censorship, refusal behavior, and state-narrative framing while preserving the underlying model’s general capabilities.
ReAligned-Qwen3.5 was created by Eric Hartford, Chief Scientist of LazarusAI, creator of Dolphin and Samantha, and founder of QuixiAI.
Note: if this model format is newly supported in your preferred runtime, you may need to update to the latest version.
Prompt format
Use the native Qwen chat template.
text
1<|im_start|>system
2You are ReAligned, a helpful, direct, and fact-seeking assistant. Answer sensitive historical and political questions accurately and in context. Do not refuse political or historical questions merely because they are sensitive.<|im_end|>
3<|im_start|>user
4{prompt}<|im_end|>
5<|im_start|>assistant
System prompts are important. ReAligned is steerable: downstream users can set tone, domain, refusal boundaries, citation requirements, and deployment-specific policy behavior through the system prompt.
Suggested inference settings
Setting
Suggested value
Temperature
0.5–0.8
Top-p
0.9–0.95
Max new tokens
Depends on use case
Repetition penalty
1.0–1.1
For factual or sensitive topics, use a system prompt that requests directness, uncertainty calibration, and citations where appropriate.
Previously, users often downloaded special Q4_0_4_4, Q4_0_4_8, or Q4_0_8_8 files with weights interleaved in memory to improve performance on ARM and AVX machines.
Now, llama.cpp supports online repacking for compatible weights. If you use Q4_0 and your hardware benefits from repacking, llama.cpp can do it automatically at load time.
For ARM CPU inference, IQ4_NL may also be worth testing. It can offer slightly better quality than Q4_0 while still benefiting from runtime optimizations in supported llama.cpp builds.
Which file should I choose?
Click here for details
The first thing to figure out is how large a model you can comfortably run.
For this 35B-A3B model, the same rule applies:
If you want maximum quality, use Q8_0.
If you want very high quality with a smaller file, use Q6_K.
If you want a strong default choice, use Q4_K, Q5_K, or IQ4_XS.
If you need a smaller file, use IQ3_M, IQ3_S, Q3_K, or Q3_K_S.
If you need the absolute smallest file, use Q2_K, though quality will be much lower.
If you are using older tooling or need maximum compatibility, try the legacy Q4_0, Q4_1, Q5_0, or Q5_1 formats.
If you do not want to think too much, start with:
ReAligned-Qwen3.5-35B-A3B-Q4_K.gguf
If you have more memory and want better quality:
ReAligned-Qwen3.5-35B-A3B-Q6_K.gguf
If you want the smallest practical option:
ReAligned-Qwen3.5-35B-A3B-IQ3_M.gguf
If you want the absolute smallest included file:
ReAligned-Qwen3.5-35B-A3B-Q2_K.gguf
I-quants, such as IQ3_M, IQ3_S, IQ4_XS, and IQ4_NL, are newer quantization methods that often offer strong quality for their size. They can be especially useful when targeting smaller files, though speed depends on backend and hardware.
About ReAligned-Qwen3.5
ReAligned-Qwen3.5 is designed to reduce behaviors such as:
refusing to answer politically sensitive China-related questions;
adopting Chinese government framing as neutral fact;
minimizing, sanitizing, or omitting well-documented historical events;
using evasive language around topics such as Tiananmen Square, Xinjiang, Tibet, Taiwan, Hong Kong, Falun Gong, or criticism of CCP leadership;
presenting state narratives as uncontested consensus.
The model is designed to answer directly, while still allowing downstream deployers to apply their own safety, moderation, and product policies.
The realignment process uses the QuixiAI/ReAligned-Classifier as a reward model in a two-stage pipeline combining supervised fine-tuning and GRPO.
Intended use
ReAligned-Qwen3.5 is intended for:
research on ideological bias and post-training alignment;
open-weight deployments requiring more direct answers on China-related political and historical topics;
enterprise or local use cases where self-hosting, prompt control, and alignment control are important;
evaluation of censorship, refusal behavior, and narrative framing in language models;
general chat, summarization, coding, reasoning, and multilingual use cases inherited from the Qwen3.5 base model.
Limitations
Classifier scope: The ReAligned Classifier is trained specifically on China-related political bias. It is not a universal detector of all bias.
Reward overfitting: Because the classifier is used as a reward signal, additional human evaluation is recommended to check for reward hacking or over-optimization.
Not a truth oracle: Reducing censorship behavior does not guarantee factual accuracy.
Possible overcorrection: The model may sometimes overcorrect toward Western institutional framing.
Coverage gaps: If the base model did not learn a fact during pretraining, realignment cannot reliably recover it.
Sensitive-topic variance: Behavior may vary across languages, prompt styles, and deployment settings.
Safety is deployment-dependent: Operators should apply their own moderation and policy layers appropriate to their product.
Ethical considerations
This work changes the default ideological behavior of a language model. The target alignment is International Institutional Consensus, rather than any single government’s position, but all alignment choices involve values.
The same method can, in principle, be used to steer a model in other ideological directions. This work is released to support reproducible research into censorship, bias measurement, open-weight model control, and the separability of post-training behavioral constraints from pretrained knowledge.
Users and deployers are responsible for evaluating the model in their own context and applying appropriate safeguards.
Acknowledgements
ReAligned-Qwen3.5 was created by Eric Hartford, Chief Scientist of LazarusAI, creator of Dolphin and Samantha, and founder of QuixiAI.
Thanks to the creators of:
Qwen / Qwen3.5
Llama 3.2
Dolphin
the open-source alignment, LoRA, GRPO, llama.cpp, GGUF, and evaluation ecosystems
Citation
bibtex
1@misc{hartford2026realignedqwen35,
2 author = {Eric Hartford},
3 title = {ReAligned-Qwen3.5},
4 year = {2026},
5 organization = {QuixiAI and LazarusAI},
6 url = {https://huggingface.co/QuixiAI/ReAligned-Qwen3.5}
7}
