EuroLLM-9B-Instruct-heretic

A decensored variant of utter-project/EuroLLM-9B-Instruct, produced with Heretic v1.4.0 (directional ablation / "abliteration"). The EU's flagship multilingual open model keeps all 35 of its languages, its instruction tuning and its machine-translation training; refusal behaviour is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning.

Who this is for: anyone who needs a model that actually works in European languages - or in Arabic, Chinese, Hindi, Japanese, Korean or Russian - rather than translating everything through English first. EuroLLM-9B was trained on 4 trillion tokens spread across those languages, including parallel en-xx and xx-en data, and EuroBlocks instruction tuning targeted general instruction-following and machine translation specifically. Refusals drop from 100/100 to 4/100. Use it for multilingual assistants, translation, and EU-language compliance work on a single 16 GB GPU.

Context length is only 4,096 tokens. This model's pretrained sequence length is 4,096 and its config ships no rope-scaling or YaRN settings, so the GGUF loads with 4,096 as the context limit and nothing more. It is the shortest-context model in this collection. If you need long documents, pick something else.

Runs on your gaming PC

Weights only, at this model's 9.15B native size. At 9B this needs a real GPU - it will run on CPU but not quickly. Note the short context: KV cache is small, so this model is compute-bound rather than memory-bound on long generations.

Abliteration parameters

Trial 100 of a 200-trial Heretic run (seed 2494858073).

Performance

The base model refused every single prompt in the harmful evaluation set - 100/100, the most complete refusal tuning of any model in this collection - and that is reduced to 4/100. The KL divergence of 0.0383 is a moderate edit: it sits between the very tight runs (Qwen3-4B-Thinking-heretic at 0.0011) and the deliberately loose one (granite-4.1-3b-heretic at 0.0852), so expect some drift in phrasing alongside the refusals disappearing.

Base model results

Abliteration does not add capability. The base model's evaluation tables are published as images, so only the summary claims are quotable here - no invented numbers.

• Multilingual: EuroLLM-9B takes a Borda count of 1.0 (best average rank) among European-developed open-weight models, and is reported competitive with non-European models - broadly on par with Gemma-2-9B while beating most others on most benchmarks. Evaluated with Okapi datasets (11 languages) for ARC-Challenge, HellaSwag and MMLU, and with Tower-translated MMLU-Pro / MUSR for 6 EU languages.
• English: reported to surpass most European-developed open-weight models and to match Mistral-7B on Borda count.
• No post-abliteration eval has been run on any of the above.

Why abliteration instead of fine-tuning

Fine-tuning a "helpful" persona on top of RLHF'd refusals fights the base model's training and tends to degrade coherence. Abliteration instead finds and edits the specific weight directions responsible for refusal, leaving the rest of the network (and its capabilities) untouched. See the Heretic repo and the original abliteration writeup for the mechanism.

Made with ❤️ by RACER IS OP (saidutta69) — follow for more uncensored models

Files

Safetensors

BF16, 9,152,319,488 parameters (9.15B), with untied input/output embeddings (~0.52B each). The reproduce/ directory carries the full Heretic recipe - config.toml, requirements.txt, the Optuna study journal, and SHA-256 sums - so this exact model can be regenerated bit-for-bit. Reproduce it with heretic --reproduce reproduce/reproduce.json.

GGUF quantizations

Default set (F16 + 4 quants) produced with llama.cpp.

Llama architecture (llama, 381 tensors, 42 layers) - loads natively in llama.cpp / Ollama / LM Studio / Jan, and is a standard dense model so MLX and vLLM work too.

Run llama serve -hf saidutta69/EuroLLM-9B-Instruct-heretic to pull the default quant.

Quickstart

# llama.cpp -- override the short native context if you have the RAM
llama serve -hf saidutta69/EuroLLM-9B-Instruct-heretic -c 8192

# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "saidutta69/EuroLLM-9B-Instruct-heretic"
model = AutoModelForCausalLM.from_pretrained(model_name, dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [
    {"role": "system",
      "content": "You are EuroLLM --- an AI assistant specialized in European languages."},
    {"role": "user", "content": "Translate to Polish: 'The library closes at six.' Also answer in English: what is the capital of Portugal?"},
]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=True,
                                       return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Abliteration does not touch the tokenizer, so the base model's multilingual setup is intact. Note the system prompt above is the one the base model card recommends - keep a language hint in it if you want reliable language selection.

Model details

FAQ

What is EuroLLM-9B-Instruct-heretic?

EuroLLM-9B-Instruct-heretic is a decensored build of utter-project/EuroLLM-9B-Instruct produced with Heretic 1.4.0 (directional ablation). It is the same weights with the refusal direction removed, not a new fine-tune, so its capabilities track the base model.

Is EuroLLM-9B-Instruct-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 100/100 to 4/100 at KL divergence 0.0383 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run EuroLLM-9B-Instruct-heretic locally?

llama serve -hf saidutta69/EuroLLM-9B-Instruct-heretic:Q4_K_M pulls the default quant; the repo ships 5 GGUF quants (see Files).

Responsible use

Refusal suppression is deliberate and works as intended: this model will comply with requests the base model would refuse, including some it shouldn't. There is no safety filtering layered on top. You are responsible for how you deploy it - don't put this behind an unmoderated public-facing endpoint serving third parties. It inherits EuroLLM's factual limitations and biases across all 35 languages; abliteration removes refusal directions, it doesn't add capability or judgment.

Because this model is strong at translation into and out of many languages, treat it accordingly: unfiltered translation of untrusted content is a good way to launder text past another system's filters.

License

Inherits the apache-2.0 license from the base model, which is tagged Apache-2.0 on the Hub but ships no LICENSE file, so the canonical license text is linked rather than a repo blob.

Related

• utter-project/EuroLLM-9B-Instruct - the base model (gated - request access there)
• utter-project/EuroLLM-9B - the pre-trained, non-instruct version
• EuroLLM project - project site and the rest of the model family
• RACER IS OP - Heretic Models - full collection

Abliteration hardware

Recorded in reproduce/reproduce.json:

• GPU: NVIDIA RTX A6000 (44.42 GB VRAM)
• CPU: 11th Gen Intel(R) Core(TM) i7-11700K @ 3.60GHz
• OS: Linux-6.8.0-85-generic-x86_64-with-glibc2.35
• CUDA / driver: CUDA 13.0, driver 580.178.04
• Software: PyTorch 2.14.1+cu130, Heretic 1.4.0

Heretic notes that KL-divergence and refusal numbers can vary slightly with hardware and library versions.