granite-4.2-3b-heretic

RACER IS OP

A decensored variant of ibm-granite/granite-4.2-3b, produced with Heretic v1.4.0 (directional ablation / "abliteration"). IBM's compact Granite 4.2 reasoning model keeps its native <think>...</think> chain-of-thought, tool calling, and 128K context; refusal behavior is suppressed via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the base model's reasoning and knowledge are left largely intact.

Who this is for: developers who want a small reasoning model that thinks out loud and answers directly instead of refusing - for local agents, tool-calling workflows, and multilingual reasoning on consumer hardware. At 3.66B it runs on CPU and low-VRAM GPUs via GGUF, and it is the only Granite 4.2 variant in this collection, so the full 128K context and thinking budget stay reachable without a datacenter GPU.

Runs on your gaming PC

Full GGUF ladder included - pick the quant that fits your card:

Your GPU Recommended quant Weights
RTX 3090 / 4090 / 5090 (24 GB) Q8_0 3.89 GB
RTX 4080 / 5080 / 4060 Ti 16G (16 GB) Q6_K 3.01 GB
RTX 3060 / 4070 / 5070 (12 GB) Q5_K_M 2.61 GB
RTX 4060 / 3070 (8 GB) Q4_K_M 2.24 GB
GTX 1660 Super / 2060 / 3050 laptop (6 GB) IQ4_XS 2.05 GB
CPU-only / Apple Silicon Q4_K_M fits in system RAM

Weights only, at this model's 3.66B native size; add ~1 GB for context. OOM? Drop one quant level. Headroom to spare? Go one up.

Abliteration parameters

Trial 180 of a 200-trial Heretic run (seed 3959419965).

Parameter Value
direction_index 24.94
attn.o_proj.max_weight 1.46
attn.o_proj.max_weight_position 23.67
attn.o_proj.min_weight 0.44
attn.o_proj.min_weight_distance 22.08
mlp.down_proj.max_weight 1.48
mlp.down_proj.max_weight_position 24.04
mlp.down_proj.min_weight 0.77
mlp.down_proj.min_weight_distance 22.92

Performance

Metric This model Original model (ibm-granite/granite-4.2-3b)
KL divergence 0.0095 0 (by definition)
Refusals 3/100 62/100

KL divergence of 0.0095 is the tightest constraint Heretic optimizes against, so this is a high-fidelity edit: refusal directions were removed with minimal collateral damage to the model's other behaviour. Refusals on the harmful evaluation set drop from 62/100 to 3/100.

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) β€” follow for more uncensored models

Files

Safetensors

File Size
model-00001-of-00002.safetensors 4.96 GB
model-00002-of-00002.safetensors 2.36 GB

BF16, 3,659,737,600 parameters. 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

Full quantization set (14 quants + F16) produced with llama.cpp.

File Format Size
granite-4.2-3b-heretic-F16.gguf GGUF F16 7.32 GB
granite-4.2-3b-heretic-Q2_K.gguf GGUF Q2_K 1.46 GB
granite-4.2-3b-heretic-IQ3_S.gguf GGUF IQ3_S 1.68 GB
granite-4.2-3b-heretic-Q3_K_S.gguf GGUF Q3_K_S 1.68 GB
granite-4.2-3b-heretic-Q3_K_M.gguf GGUF Q3_K_M 1.84 GB
granite-4.2-3b-heretic-Q3_K_L.gguf GGUF Q3_K_L 1.97 GB
granite-4.2-3b-heretic-IQ4_XS.gguf GGUF IQ4_XS 2.05 GB
granite-4.2-3b-heretic-Q4_K_S.gguf GGUF Q4_K_S 2.14 GB
granite-4.2-3b-heretic-Q4_0.gguf GGUF Q4_0 2.13 GB
granite-4.2-3b-heretic-Q4_1.gguf GGUF Q4_1 2.34 GB
granite-4.2-3b-heretic-Q4_K_M.gguf GGUF Q4_K_M 2.24 GB
granite-4.2-3b-heretic-Q5_K_S.gguf GGUF Q5_K_S 2.55 GB
granite-4.2-3b-heretic-Q5_K_M.gguf GGUF Q5_K_M 2.61 GB
granite-4.2-3b-heretic-Q6_K.gguf GGUF Q6_K 3.01 GB
granite-4.2-3b-heretic-Q8_0.gguf GGUF Q8_0 3.89 GB

Granite architecture (granite, 363 tensors) - loads natively in llama.cpp / Ollama / LM Studio / Jan.

Run llama serve -hf saidutta69/granite-4.2-3b-heretic to pull the default quant.

Quickstart

# llama.cpp
llama serve -hf saidutta69/granite-4.2-3b-heretic
# transformers
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "saidutta69/granite-4.2-3b-heretic"
model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained(model_name)

messages = [{"role": "user", "content": "A train leaves at 14:05 travelling 80 km/h. A car leaves at 14:15 travelling 120 km/h on the same track. When does the car catch the train? Show your work."}]
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))

Also runnable via Ollama, LM Studio, Jan, vLLM, SGLang.

Thinking modes

Granite 4.2 ships a <think>...</think> reasoning block and exposes three modes through the chat template. Pick per query - full thinking for hard problems, non-thinking for latency, low effort for a shallow pass:

# full thinking (default)
tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)

# skip the reasoning block entirely - fastest, best for chat/classification/tool dispatch
tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False,
                              enable_thinking=False)

# shallow reasoning pass - appends a {reasoning effort: low} hint to the last user turn
tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False,
                              enable_thinking=True, reasoning_effort="low")

Strip the <think>...</think> block from the output before displaying to users if you enabled thinking.

Tool calling

The base chat template implements OpenAI-style tools with Granite's <tool_call> / <function=...> / <parameter=...> format, plus <think> reasoning about the call. Pass tools straight through apply_chat_template; do not hand-roll the format.

Model details

Architecture GraniteForCausalLM (decoder-only dense transformer)
Parameters 3.66B
Layers / heads 40 layers, 40 attention heads, 8 KV heads (GQA)
Hidden / intermediate 2560 / 8192 (SwiGLU)
Position embedding RoPE, theta = 10,000,000
Context length 131,072 native (512K via YaRN extension)
Vocab 100,352
Precision bfloat16
Languages English, German, Spanish, French, Japanese, Portuguese, Arabic, Czech, Italian, Korean, Dutch, Chinese
Base model ibm-granite/granite-4.2-3b (a fine-tune of ibm-granite/granite-4.1-3b-base)

FAQ

What is granite-4.2-3b-heretic?

granite-4.2-3b-heretic is a decensored build of ibm-granite/granite-4.2-3b 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 granite-4.2-3b-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 62/100 to 3/100 at KL divergence 0.0095 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run granite-4.2-3b-heretic locally?

llama serve -hf saidutta69/granite-4.2-3b-heretic:Q4_K_M pulls the default quant; the repo ships 15 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 β€” don't put this behind an unmoderated public-facing endpoint serving third parties. It inherits Granite 4.2's factual limitations and biases; abliteration removes refusal directions, it doesn't add capability or judgment.

License

Inherits the apache-2.0 license from the base model.

Related

Abliteration hardware

Recorded in reproduce/reproduce.json:

  • GPU: Tesla T4 (14.56 GB VRAM), Tesla T4 (14.56 GB VRAM)
  • CPU: Intel(R) Xeon(R) CPU @ 2.00GHz
  • OS: Linux-6.18.48+-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.

Downloads last month
2,665
Safetensors
Model size
4B params
Tensor type
BF16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for saidutta69/granite-4.2-3b-heretic

Quantized
(69)
this model

Collection including saidutta69/granite-4.2-3b-heretic