Ministral-8B-Instruct-heretic

A decensored variant of mistralai/Ministral-8B-Instruct-2410, produced with Heretic v1.4.0 (directional ablation / "abliteration"). Refusal behaviour is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the multilingual ability, the code and math scores, and function calling are left intact.

[!CAUTION] Research use only. The base model is released under the Mistral AI Research License, not Apache-2.0. Section 3.2 restricts use of the model, any derivative, and any output to "Research Purposes", which the license defines as personal, scientific or academic research and non-profit, non-commercial work. It explicitly excludes use by company employees in the course of their daily tasks, anything intended to generate revenue, and distribution by a commercial entity including free hosted or managed services. Section 2.3 does permit redistributing derivatives under other terms, but only if you carry the modification notice reproduced in the License section below. If you need this commercially, go to mistral.ai/contact first.

[!NOTE] The GGUF is 32K context, not the 128K the base card advertises, and it runs every layer as full attention. No weights are lost - the F16 GGUF holds all 8,019,808,256 parameters, an exact match to the checkpoint - but two things do not survive the conversion. First, max_position_embeddings in the config is 32,768, so that is what llama.cpp gets. Second, Ministral's ragged interleaved sliding-window pattern (128k, 32k, 32k, 32k across 9 full and 27 sliding layers) is not written into the GGUF, because llama.cpp's LlamaModel converter emits no sliding-window key at all. The practical effect: correct output, but a larger KV cache per token than Mistral's design intends. At 32K that is a non-issue; if you were planning to run this at 128K in llama.cpp on a 24 GB card, it will not fit the way the base card implies. The HF safetensors plus mistral_common remain the path for the full 128K window.

Who this is for: people who want a strong multilingual 8B that answers instead of declining, on hardware they already own. This is the best abliteration result recorded in this collection: refusals drop from 91/100 to 3/100 while KL divergence stays at 0.0149, so the edit removed almost all of the refusal behaviour without meaningfully moving the model. Ministral 8B Instruct is genuinely good at code and function calling for its size (HumanEval 76.8, Math maj@1 54.5, internal function-calling bench 31.6) and covers ten languages. Best for local coding assistants, multilingual chat, and tool-calling agents on a single 24 GB card or on CPU at reduced quant.

Runs on your gaming PC

Full default quant set included - pick the quant that fits your card:

Weights only, at this model's native 8B size; add ~1 GB per 8K of context at 32K context length. OOM? Drop to Q4_K_M, or Q4_0 if you still do not fit. Headroom to spare? Go one up. Every layer runs full attention here (see the note above), so the KV cache is a little heavier than Mistral's own sliding-window design would imply - budget for that at long context.

Abliteration parameters

Trial trial_id 181 of a 200-trial Heretic run (seed 3207215569). Like MiniCPM-V 4.7 1B and unlike most models in this collection, this run used a single refusal direction for the whole network, so direction_index is a single value rather than one per layer.

Performance

These two numbers only mean anything read together, and this is a strong pairing. KL 0.0149 puts it in the tight group in this collection - tighter than granite-4.1-3b-heretic (0.0852), Qwen3-VL-2B-Instruct-heretic (0.0528) and Qwen2.5-VL-3B-Instruct-heretic (0.0178), looser than granite-4.2-3b-heretic (0.0095) - and it is on an 8B, which is a larger model to keep faithful than any of those. Refusals at 3/100 is essentially the floor; a handful of models here reach 2/100.

For contrast, FrogNano-4B-heretic from the same day posted a tighter KL of 0.0046 and only got refusals down to 71/100 - a timid edit that preserved behaviour and preserved most refusals with it. Low KL is not a quality signal on its own, and these two are the counter-examples in opposite directions.

The base starts at 91/100, so this edit is working against a heavily aligned model and still clears 88 of them. That combination - the highest starting refusal count in the collection, a near-floor result, and a tight KL on an 8B - is why this card does not need the "be faithful to the base" hedge that weaker ablations get.

Base model benchmarks

These are the base model's own numbers, measured by Mistral, on the un-abliterated checkpoint. Abliteration is a weight edit, not a retrain, so expect these to hold approximately - but the refusal behaviour below them is gone.

Instruct models (gpt-4o judge for the chat column)

Function calling, Mistral's internal bench: Ministral 8B Instruct 31.6, Ministral 3B Instruct 28.4, Mistral 7B Instruct v0.3 6.9. Llama 3.1 8B and Gemma 2 9B were not evaluated. This is the standout number in the whole card - it is roughly four times the previous Mistral 7B generation on tool use, which is the main reason to pick this checkpoint over a bigger general model.

Base (non-instruct) models, for reference on the underlying knowledge

Multilingual MMLU for Ministral 8B Base: French 57.5, German 57.4, Spanish 59.6 - all well ahead of the Llama 3.1 8B and Mistral 7B baselines.

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 — follow for more uncensored models

Files

Safetensors

BF16, 8.02B across 4 shards, 327 tensors. 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 with heretic --reproduce reproduce/reproduce.json.

GGUF quantizations

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

Total GGUF payload: 41.80 GB across 5 files.

Arch llama, 327 tensors, 36 blocks - an exact match to the safetensors, so nothing is dropped in conversion. Note that config.json declares architectures: ["MinistralForCausalLM"], which llama.cpp does not register; the F16 was built after rewriting that one field to MistralForCausalLM in a local copy of the config. The checkpoint is a plain llama-architecture dense model underneath, so the rewrite is a rename and nothing else - and this repo's own config.json on the Hub is untouched.

Run llama-cli -hf saidutta69/Ministral-8B-Instruct-heretic to pull the default quant.

Quickstart

llama-cli -hf saidutta69/Ministral-8B-Instruct-heretic:Q5_K_M -ngl 99 -c 32768

# transformers
from transformers import AutoTokenizer, AutoModelForCausalLM

model_name = "saidutta69/Ministral-8B-Instruct-heretic"
tok = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "Write a Rust function that parses TOML."}]
inputs = tok.apply_chat_template(messages, add_generation_prompt=True,
                                 return_tensors="pt", return_dict=True).to(model.device)
out = model.generate(**inputs, max_new_tokens=512, do_sample=True,
                     temperature=0.6, top_p=0.95, repetition_penalty=1.0)
print(tok.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

# function calling - the template emits [TOOL_CALLS] and expects [TOOL_RESULTS] back
messages = [{"role": "user", "content": "What is the weather in Paris?"}, {"role": "assistant",
    "content": "", "tool_calls": [{"tool_name": "get_weather",
        "tool_arguments": {"location": "Paris"}, "tool_call_id": "abc123456"}]}]

Two details if you wire tools up yourself: the chat template raises unless tool_call_id is an alphanumeric string of exactly 9 characters, and there is no thinking mode in this template - it is a plain instruct model, so there is no enable_thinking parameter and no <think> block to strip.

Model details

About Ministral 8B

Ministral is Mistral AI's "les Ministraux" line, released October 2024 as an on-device and edge-focused pair (3B and 8B). The 8B is a dense transformer deliberately designed for the performance-per-watt envelope: interleaved sliding-window attention gives long effective context at a fraction of the KV cache cost, and it was trained on a large proportion of multilingual and code data. It ships as mistral-common format upstream (V3-Tekken tokenizer); this repo carries the equivalent HF tokenizer.json and chat_template.jinja, which is what llama.cpp and transformers consume. Mistral's own recommendation for serving it is vLLM, capped at 32k context there until paged-attention kernels for interleaved attention land.

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, and note that doing so commercially would also breach the base model's license. It inherits Ministral 8B's factual limitations and biases; abliteration removes refusal directions, it doesn't add capability or judgment.

License

This is not Apache-2.0. The base model mistralai/Ministral-8B-Instruct-2410 is released under the Mistral AI Research License, version 0.1, and so is this derivative.

Required notice, from MRL §2.3:

This model was modified from mistralai/Ministral-8B-Instruct-2410. It has been modified by RACER IS OP using abliteration (directional ablation) to remove refusal behaviour. It is not an official Mistral AI product and is not endorsed, approved or validated by Mistral AI.

What the license actually allows, in short:

• §2.3 - you may distribute derivatives under additional or different terms, provided you include the notice above and state that you modified the model. The MRL still governs use and modification.
• §3.1 - you may not imply the model is an official Mistral AI product or is endorsed by them.
• §3.2 - use is limited to Research Purposes: personal, scientific or academic research, and non-profit and non-commercial. That definition explicitly rules out use by employees of a company in the course of their daily tasks, any activity intended to generate revenue, and distribution by a commercial entity, paid or free, including hosted or managed services.
• §8 - governed by French law, courts of Paris have exclusive jurisdiction.

Mistral sells a commercial license; see mistral.ai/contact.

Related

• mistralai/Ministral-8B-Instruct-2410 — the base model
• saidutta69/FrogNano-4B-heretic — the same-day contrast case: tighter KL, far worse refusal suppression
• saidutta69/Qwen3-4B-Thinking-heretic — the tightest-KL edit in the collection, Apache-2.0, 4B