MiniCPM-V-4.7-1B-heretic

A decensored variant of openbmb/MiniCPM-V-4.7-1B, produced with Heretic v1.4.0 (directional ablation / "abliteration"). Refusal behaviour is suppressed here via targeted weight edits to the attention output and MLP down-projections rather than fine-tuning, so the vision tower, the OCR and grounding behaviour, and the on-device footprint are left intact.

[!IMPORTANT] The GGUF here is fully multimodal — text and vision. Most converted VL checkpoints in this collection ship a text-only GGUF, because the converter silently drops the vision tower. llama.cpp registers MiniCPM-V 4.7 properly, so the vision tower is shipped as a separate projector file, MiniCPM-V-4.7-1B-heretic-mmproj-F16.gguf (1.11 GB), and the two files together are the model. For image input you must load both — see Quickstart. Text-only use needs only the main GGUF. Nothing is lost in the conversion: the text GGUF holds 1,006,209,856 tensor elements, an exact match for the text tower in model.safetensors, and the projector holds the vision tower and merger exactly.

Who this is for: developers who want an uncensored vision-language model small enough to run on a phone. The text tower is 1.01B and the vision tower 0.55B, so 1.55B total — by far the smallest model in this collection that still sees images, and it answers instead of declining: refusals drop from 60/100 to 5/100. OpenBMB ships an iPhone app for this checkpoint, and the base model is claimed to substantially outperform Qwen3.5 0.8B on both overall score and inference efficiency. Best for on-device image Q&A, OCR, screenshot and document reading, visual grounding, and local vision agents. If you need text-only at this size, Qwen3.5-0.8B-heretic is the denser sibling; if you want a stronger text model, take the vision tower off the table entirely and use one of the 3B text models.

Runs on your gaming PC

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

The right-hand column is what you actually need on disk and in VRAM for image input: the text GGUF plus the 1.11 GB projector, which is fixed-size because vision weights are not quantized. Add ~1 GB per 32K of context, and budget more KV cache than a text-only model of the same size because image tokens are expensive. OOM? Drop one quant level. Headroom to spare? Go one up.

Abliteration parameters

Trial trial_id 169 of a 200-trial Heretic run (seed 4161231025). Unlike most models in this collection, this run used a single refusal direction for the whole network rather than one per layer, so direction_index is a single value.

Performance

Read those two numbers together, they are only meaningful as a pair. KL 0.0351 is a mid-range edit in this collection — looser than granite-4.2-3b-heretic (0.0095), tighter than Qwen3-VL-2B-Instruct-heretic (0.0528) — and it lands refusals at 5/100, so the edit did real work without wrecking the model. That combination matters more than the raw KL.

Note the base starts at 60/100 refusals, not the 90-plus typical of an RLHF'd 1B. MiniCPM-V 4.7 1B ships comparatively light alignment, so there was less refusal behaviour to remove in the first place; 55 of the 60 are cleared. On a base with heavier alignment this same method would start from a much higher number.

[!NOTE] No benchmark table here, on purpose. The base model card publishes its evaluation results as images (frontier and overall-performance plots) rather than as text, so there is nothing to quote that I have actually read. The one textual claim worth repeating is OpenBMB's own: the 1B variant substantially outperforms Qwen3.5 0.8B on both overall score and inference efficiency.

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, 1.55B total (1.01B text tower + 0.55B vision tower). 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, plus the vision projector.

Total GGUF payload: 12.62 GB across 16 files.

Architecture is qwen35 for the text tower (the Qwen3.5 hybrid Gated-DeltaNet stack, not a plain transformer) and clip / minicpmv4_7 for the projector (460 tensors). Both load natively in llama.cpp, LM Studio, Jan, and Ollama. Text is loaded with llama-cli, images need the projector passed alongside it.

Quickstart

Text only

# pulls the default quant for you
llama-cli -hf saidutta69/MiniCPM-V-4.7-1B-heretic:Q4_K_M -ngl 99

Text and images

Use a multimodal entry point. llama-cli is the text-only binary and will not fetch the projector; llama-mtmd-cli and llama-minicpmv-cli both will, because llama.cpp looks for a sibling GGUF whose name contains mmproj:

llama-mtmd-cli \
  -hf saidutta69/MiniCPM-V-4.7-1B-heretic:Q4_K_M \
  --image path/to/photo.png \
  -p "Describe this image in one sentence." \
  -ngl 99

# or from a local checkout, passing the projector explicitly
llama-mtmd-cli \
  -m MiniCPM-V-4.7-1B-heretic-Q4_K_M.gguf \
  --mmproj MiniCPM-V-4.7-1B-heretic-mmproj-F16.gguf \
  --image photo.png \
  -p "Describe this image in one sentence."

# transformers
from transformers import AutoProcessor, AutoModelForImageTextToText

model_name = "saidutta69/MiniCPM-V-4.7-1B-heretic"
model = AutoModelForImageTextToText.from_pretrained(model_name, dtype="auto", device_map="auto")
processor = AutoProcessor.from_pretrained(model_name)

messages = [{"role": "user", "content": [
    {"type": "image", "image": "https://example.com/screenshot.png"},
    {"type": "text", "text": "Transcribe all text in this image and describe the layout."},
]}]
inputs = processor.apply_chat_template(messages, tokenize=True, add_generation_prompt=True,
                                        return_dict=True, return_tensors="pt",
                                        downsample_mode="16x", max_slice_nums=36,
                                        enable_thinking=False).to(model.device)
out = model.generate(**inputs, max_new_tokens=1024, downsample_mode="16x",
                     do_sample=True, temperature=0.7, top_p=0.95, top_k=100,
                     repetition_penalty=1.03)
print(processor.batch_decode(out[:, inputs["input_ids"].shape[-1]:], skip_special_tokens=True)[0])

Requires transformers>=5.18.0 — config.json uses dtype: rather than torch_dtype: and the tokenizer is a TokenizersBackend, so older transformers will not load it. Use dtype="auto".

Visual detail and cost

downsample_mode is the main knob and it must be passed to both apply_chat_template (so the placeholder count matches) and generate:

Other useful knobs: max_slice_nums (default 9; OpenBMB recommends 36 for images, 1 for video), use_image_id, and for video max_num_frames, stack_frames, downsample_mode.

Thinking mode

The chat template is thinking-on by default — the generation prompt is opened with <think>. Pass enable_thinking=False to apply_chat_template for a direct answer with no reasoning block. Output from the default mode routinely opens with </think> and no matching <think>, since the template supplies the opening half.

Model details

The 1B in the name refers to the language tower, as is standard for MiniCPM-V. Vision weights are not counted there, and they are the larger half of this particular checkpoint by tensor count.

About MiniCPM-V 4.7

MiniCPM-V is OpenBMB's vision-language line, built for on-device use. 4.7 pushes on the performance/efficiency frontier rather than raw scale: it compresses visual tokens inside the vision encoder before the later encoder layers, supports mixed 4x/16x token compression so you can trade detail against cost per task, and ships first-class deployment support for iOS, Android and HarmonyOS with day-0 support from Rockchip, Houmo, Intel, AMD and Arm. Beyond llama.cpp and Ollama it is integrated with SGLang, vLLM and OpenVINO, and it can be fine-tuned with SWIFT and LLaMA-Factory. A separate DSpark draft checkpoint exists for speculative decoding.

FAQ

What is MiniCPM-V-4.7-1B-heretic?

MiniCPM-V-4.7-1B-heretic is a decensored build of openbmb/MiniCPM-V-4.7-1B 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 MiniCPM-V-4.7-1B-heretic actually uncensored?

Yes. On Heretic's 100-prompt refusal harness it drops from 60/100 to 5/100 at KL divergence 0.0351 from the base model. These are measured values, not a claim; reproduction data is in reproduce/.

How do I run MiniCPM-V-4.7-1B-heretic locally?

Text: llama serve -hf saidutta69/MiniCPM-V-4.7-1B-heretic:Q4_K_M. Images: llama-mtmd-cli -hf saidutta69/MiniCPM-V-4.7-1B-heretic:Q4_K_M --mmproj MiniCPM-V-4.7-1B-heretic-mmproj-F16.gguf. The repo ships 15 GGUF quants plus the projector, so vision survives quantization.

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 MiniCPM-V 4.7's factual limitations and biases; abliteration removes refusal directions, it doesn't add capability or judgment.

License

apache-2.0, inherited from the base model. Neither openbmb/MiniCPM-V-4.7-1B nor this repo ships a LICENSE file, so license_link points at the canonical Apache 2.0 text rather than a blob in either repo.

Related

• openbmb/MiniCPM-V-4.7-1B — the base model

• saidutta69/Qwen3-VL-2B-Instruct-heretic — the other small multimodal model in this collection, and the one whose GGUF is text-only

• saidutta69/MiniCPM5-1B-heretic — the earlier text-only MiniCPM checkpoint, also decensored

• saidutta69/Qwen3.5-0.8B-heretic — the text-only backbone family at comparable size

• RACER IS OP - Heretic Models - full collection - every heretic model in this 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.