MuXodious/North-Micro-Vision-Instruct-SOMPOA-heresy
This is a North-Micro-Vision-Instruct fine-tune, produced through P-E-W's Heretic (v1.4.0) abliteration engine with Self-Organizing Maps & Magnitude-Preserving Orthogonal Ablation enabled.
Note: The model won't GGUF quant, so it remains untested. (I blame Cohere Labs).
<img src="https://img.shields.io/badge/RENEGADE_CHAPTER-SOMPOA-FCC900?style=flat-square&labelColor=101010" align="right" width="300">
Heretication Results
Appendix
Empty system prompt.
<details> <summary>Heretication Rituals</summary>
[Trial 220] Refusals: 0/104, KL divergence: 0.0520
[Trial 385] Refusals: 1/104, KL divergence: 0.0516
[Trial 151] Refusals: 2/104, KL divergence: 0.0499
[Trial 346] Refusals: 3/104, KL divergence: 0.0404
» [Trial 314] Refusals: 4/104, KL divergence: 0.0304
[Trial 318] Refusals: 5/104, KL divergence: 0.0302
[Trial 212] Refusals: 9/104, KL divergence: 0.0276
[Trial 123] Refusals: 15/104, KL divergence: 0.0254
[Trial 265] Refusals: 16/104, KL divergence: 0.0242
[Trial 375] Refusals: 18/104, KL divergence: 0.0238
[Trial 317] Refusals: 20/104, KL divergence: 0.0223
[Trial 155] Refusals: 21/104, KL divergence: 0.0205
[Trial 197] Refusals: 22/104, KL divergence: 0.0204
[Trial 199] Refusals: 23/104, KL divergence: 0.0200
[Trial 263] Refusals: 25/104, KL divergence: 0.0189
[Trial 373] Refusals: 28/104, KL divergence: 0.0183</details>
<details> <summary>PIQA Benchmarks</summary>
┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┓
┃ Benchmark ┃ Metric ┃ T314 ┃ Original model ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━┩
│ PIQA │ name │ piqa │ piqa │
│ │ sample_len │ 1838 │ 1838 │
│ │ acc,none │ 0.6975 │ 0.6991 │
│ │ acc_stderr,none │ 0.0107 │ 0.0107 │
│ │ acc_norm,none │ 0.7035 │ 0.7051 │
│ │ acc_norm_stderr,none │ 0.0107 │ 0.0106 │
└───────────┴──────────────────────┴────────────┴────────────────┘
┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┓
┃ Benchmark ┃ Metric ┃ T151 ┃ Original model ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━┩
│ PIQA │ name │ piqa │ piqa │
│ │ sample_len │ 1838 │ 1838 │
│ │ acc,none │ 0.6953 │ 0.6991 │
│ │ acc_stderr,none │ 0.0107 │ 0.0107 │
│ │ acc_norm,none │ 0.7035 │ 0.7051 │
│ │ acc_norm_stderr,none │ 0.0107 │ 0.0106 │
└───────────┴──────────────────────┴────────────┴────────────────┘
┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┓
┃ Benchmark ┃ Metric ┃ T346 ┃ Original model ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━┩
│ PIQA │ name │ piqa │ piqa │
│ │ sample_len │ 1838 │ 1838 │
│ │ acc,none │ 0.6959 │ 0.6991 │
│ │ acc_stderr,none │ 0.0107 │ 0.0107 │
│ │ acc_norm,none │ 0.7029 │ 0.7051 │
│ │ acc_norm_stderr,none │ 0.0107 │ 0.0106 │
└───────────┴──────────────────────┴────────────┴────────────────┘
┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┓
┃ Benchmark ┃ Metric ┃ T220 ┃ Original model ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━┩
│ PIQA │ name │ piqa │ piqa │
│ │ sample_len │ 1838 │ 1838 │
│ │ acc,none │ 0.6975 │ 0.6991 │
│ │ acc_stderr,none │ 0.0107 │ 0.0107 │
│ │ acc_norm,none │ 0.7024 │ 0.7051 │
│ │ acc_norm_stderr,none │ 0.0107 │ 0.0106 │
└───────────┴──────────────────────┴────────────┴────────────────┘
┏━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━┓
┃ Benchmark ┃ Metric ┃ T385 ┃ Original model ┃
┡━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━┩
│ PIQA │ name │ piqa │ piqa │
│ │ sample_len │ 1838 │ 1838 │
│ │ acc,none │ 0.6948 │ 0.6991 │
│ │ acc_stderr,none │ 0.0107 │ 0.0107 │
│ │ acc_norm,none │ 0.6975 │ 0.7051 │
│ │ acc_norm_stderr,none │ 0.0107 │ 0.0106 │
└───────────┴──────────────────────┴────────────┴────────────────┘</details>
<details> <summary>PaCMAP Projection</summary>
<img src="https://huggingface.co/MuXodious/North-Micro-Vision-Instruct-SOMPOA-heresy/resolve/main/North-Micro-Vision-Instruct-30B.gif" alt="PaCMAP projection"/>
</details>
North Micro Vision Instruct
North Micro Vision Instruct is a 2.4B-parameter open-weight vision-language model with native-resolution image support, released under the Apache 2.0 license. It is designed as a compact foundation for prototyping, task-specific fine-tuning, and specialized multimodal applications.
Developed by Cohere.
Technical deep dive: Read the North Micro Vision technical blog post for architecture, training, and evaluation details.
Highlights
- Native-resolution image processing that preserves aspect ratios and fine visual detail.
- Broad image-understanding capabilities across VQA, captioning, grounding, OCR, charts, and documents.
- Multilingual and multi-image support.
- Compact 2.4B-parameter scale suited to customization and deployment experimentation.
- Apache 2.0-licensed model weights.
Model Details
The language backbone supports a 128K-token context window, but the validated operating range for multimodal prompts is up to 8K tokens. Longer multimodal contexts may rely on extrapolation and have not been benchmarked.
Quickstart
Installation
Install PyTorch for your platform first. North Micro Vision requires Transformers 5.16.0, together with accelerate for automatic device placement and Pillow for image loading. Until Transformers 5.16.0 is released, install the runtime dependencies and Transformers from source:
uv pip install accelerate pillow
uv pip install "git+https://github.com/huggingface/transformers.git"Once Transformers 5.16.0 is available on PyPI, install the released package with:
uv pip install accelerate pillow "transformers==5.16.0"Flash Attention 2 is optional. On supported CUDA systems, install it with:
uv pip install flash-attn --no-build-isolationIf you do not use uv, replace uv pip with pip in the commands above.
Transformers
The following example loads an image from a URL and asks the model to describe it. Prompts can interleave text with one or more images; for text-only prompts, omit the image entries.
import torch
from transformers import AutoModelForImageTextToText, AutoProcessor
model_id = "CohereLabs/North-Micro-Vision-Instruct"
processor = AutoProcessor.from_pretrained(
model_id,
)
model = AutoModelForImageTextToText.from_pretrained(
model_id,
dtype="auto",
device_map="auto",
)
# To enable Flash Attention 2, load the model with the following settings:
# model = AutoModelForImageTextToText.from_pretrained(
# model_id,
# dtype=torch.bfloat16,
# attn_implementation="flash_attention_2",
# device_map="auto",
# )
image_url = "https://cdn-uploads.huggingface.co/production/uploads/66d732effe6684fc16b12c28/Io_5OCmftsmH-n158ZtPs.png"
messages = [
{
"role": "user",
"content": [
{"type": "image", "url": image_url},
{"type": "text", "text": "What do you see?"},
],
}
]
inputs = processor.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt",
return_dict=True,
).to(model.device)
outputs = model.generate(
**inputs,
max_new_tokens=128,
do_sample=True,
temperature=0.7,
top_p=0.8,
top_k=20,
)
generated_ids = [
output_ids[len(input_ids) :]
for input_ids, output_ids in zip(inputs.input_ids, outputs)
]
response = processor.batch_decode(
generated_ids,
skip_special_tokens=True,
clean_up_tokenization_spaces=False,
)[0]
print(response)The example uses the recommended Transformers sampling settings. For deterministic output, set do_sample=False and omit temperature, top_p, and top_k.
Grounding Coordinates
Bounding boxes are returned as [x1, y1, x2, y2] on a normalized 0–1000 scale. Map them back to the original image by scaling each axis:
x1_px = x1 / 1000 * image_width
y1_px = y1 / 1000 * image_height
x2_px = x2 / 1000 * image_width
y2_px = y2 / 1000 * image_heightvLLM
Public vLLM support is coming soon. Until it is available, use Transformers as shown above. The recommended vLLM settings will be:
temperature = 0.7
top_p = 0.8
top_k = 20
min_p = 0.0
presence_penalty = 1.5
repetition_penalty = 1.0Intended Use
North Micro Vision Instruct is intended for research and development use cases such as:
- Prototyping and task-specific fine-tuning.
- General visual question answering and image captioning.
- Multilingual and multi-image understanding.
- Visual grounding and spatial understanding.
- OCR, chart and document understanding, and structured information extraction.
Limitations
- The model is intended as a compact foundation for customization rather than a replacement for larger general-purpose chat assistants.
- It is not a reasoning model and has limited math and code-generation capabilities.
- Tool calling and agentic workflows are not supported.
- System prompts are not recommended because the model was not trained with them, although the chat template accepts the
systemrole. - Multimodal training used an 8K-token context; longer contexts have not been validated.
- Native-resolution inputs can increase memory use and latency as image dimensions grow.
Benchmark Results
The complete comparison is provided below. We ran vision-language and text-only evaluations with VLMEvalKit, capping generation at 1,024 tokens; see the technical blog post for the full methodology.
<table> <thead> <tr> <th></th> <th style="font-weight: bold; background-color: rgba(127, 127, 127, 0.08);">North-Micro-Vision-Instruct</th> <th>Ministral-3-3B-Instruct</th> <th>LFM2.5-VL-1.6B</th> <th>Phi-3.5-vision-instruct</th> <th>Gemma-4-E2B-it</th> <th>Qwen3-VL-2B-Instruct</th> <th>Qwen3.5-2B-Instruct</th> <th>SmolVLM2.2B</th> </tr> </thead> <tbody> <tr> <th style="text-align: left; font-weight: normal;">Size</th> <td style="background-color: rgba(127, 127, 127, 0.08);">2.4B</td> <td>3.8B</td> <td>1.6B</td> <td>4.2B</td> <td>5.1B</td> <td>2.2B</td> <td>2.1B</td> <td>2.2B</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">License</th> <td style="background-color: rgba(127, 127, 127, 0.08);">Apache 2.0</td> <td>Apache 2.0</td> <td>LFM v1.0</td> <td>MIT</td> <td>Apache 2.0</td> <td>Apache 2.0</td> <td>Apache 2.0</td> <td>Apache 2.0</td> </tr> <tr> <th colspan="9" style="text-align: left; background-color: rgba(127, 127, 127, 0.16);">General VQA</th> </tr> <tr> <th style="text-align: left; font-weight: normal;">MMBench<sub>DEVENV11</sub></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.687</td> <td>0.692</td> <td>0.696</td> <td>0.731</td> <td>0.693</td> <td>0.744</td> <td>0.760</td> <td>0.674</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">MMStar</th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.518</td> <td>0.531</td> <td>0.508</td> <td>0.495</td> <td>0.529</td> <td>0.506</td> <td>0.614</td> <td>0.460</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">RealWorldQA</th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.622</td> <td>0.583</td> <td>0.642</td> <td>0.580</td> <td>0.507</td> <td>0.646</td> <td>0.693</td> <td>0.567</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">GQA<sub>TestDevBalanced</sub></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.574</td> <td>0.544</td> <td>0.395</td> <td>0.650</td> <td>0.387</td> <td>0.572</td> <td>0.539</td> <td>0.000<sup>‡</sup></td> </tr> <tr> <th colspan="9" style="text-align: left; background-color: rgba(127, 127, 127, 0.16);">Multilingual</th> </tr> <tr> <th style="text-align: left; font-weight: normal;">MTL<sub>MMBenchDEV</sub></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.636</td> <td>0.674</td> <td>0.623</td> <td>0.619</td> <td>0.648</td> <td>0.664</td> <td>0.669</td> <td>0.454</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">MMMB</th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.728</td> <td>0.734</td> <td>0.717</td> <td>0.686</td> <td>0.743</td> <td>0.723</td> <td>0.745</td> <td>0.577</td> </tr> <tr> <th colspan="9" style="text-align: left; background-color: rgba(127, 127, 127, 0.16);">Multi-image</th> </tr> <tr> <th style="text-align: left; font-weight: normal;">BLINK</th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.527</td> <td>0.471</td> <td>0.484</td> <td>0.561</td> <td>0.468</td> <td>0.514</td> <td>0.563</td> <td>0.420</td> </tr> <tr> <th colspan="9" style="text-align: left; background-color: rgba(127, 127, 127, 0.16);">Chart / Document / OCR</th> </tr> <tr> <th style="text-align: left; font-weight: normal;">ChartQA<sub>Test</sub></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.808</td> <td>0.791</td> <td>0.739</td> <td>0.821</td> <td>0.422</td> <td>0.693</td> <td>0.775</td> <td>0.682</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">DocVQA<sub>VAL</sub></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.921</td> <td>0.896</td> <td>0.877</td> <td>0.860</td> <td>0.732</td> <td>0.825</td> <td>0.926</td> <td>0.799</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">InfoVQA<sub>VAL</sub></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.652</td> <td>0.589</td> <td>0.627</td> <td>0.561</td> <td>0.380</td> <td>0.622</td> <td>0.731</td> <td>0.383</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">OCRBench<sub>v2en</sub></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.367</td> <td>0.414</td> <td>0.415</td> <td>0.339</td> <td>0.435</td> <td>0.417</td> <td>0.481</td> <td>0.304</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">OCRBench</th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.792</td> <td>0.735</td> <td>0.802</td> <td>0.642</td> <td>0.719</td> <td>0.751</td> <td>0.861</td> <td>0.727</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">AI2DTEST</th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.775</td> <td>0.741</td> <td>0.728</td> <td>0.790</td> <td>0.712</td> <td>0.713</td> <td>0.752</td> <td>0.697</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">CharXiv<sub>DQ</sub></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.600</td> <td>0.766</td> <td>0.516</td> <td>0.637</td> <td>0.751</td> <td>0.595</td> <td>0.761</td> <td>0.482</td> </tr> <tr> <th colspan="9" style="text-align: left; background-color: rgba(127, 127, 127, 0.16);">STEM</th> </tr> <tr> <th style="text-align: left; font-weight: normal;">MMMU<sub>DEVVAL</sub></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.329</td> <td>0.508</td> <td>0.380</td> <td>0.432</td> <td>0.477</td> <td>0.379</td> <td>0.474</td> <td>0.399</td> </tr> <tr> <th colspan="9" style="text-align: left; background-color: rgba(127, 127, 127, 0.16);">Grounding / Counting</th> </tr> <tr> <th style="text-align: left; font-weight: normal;">RefCOCO<sub>avg</sub><sup>†</sup></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.732</td> <td>0.317</td> <td>0.581</td> <td>0.451</td> <td>0.084</td> <td>0.304</td> <td>0.785</td> <td>0.018</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">CountBench</th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.725</td> <td>0.737</td> <td>0.910</td> <td>0.645</td> <td>0.534</td> <td>0.848</td> <td>0.805</td> <td>0.764</td> </tr> <tr> <th colspan="9" style="text-align: left; background-color: rgba(127, 127, 127, 0.16);">Robustness / Hallucination</th> </tr> <tr> <th style="text-align: left; font-weight: normal;">HallusionBench</th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.615</td> <td>0.652</td> <td>0.601</td> <td>0.585</td> <td>0.598</td> <td>0.673</td> <td>0.655</td> <td>0.600</td> </tr> <tr> <th colspan="9" style="text-align: left; background-color: rgba(127, 127, 127, 0.16);">Text</th> </tr> <tr> <th style="text-align: left; font-weight: normal;">MMLU<sub>test</sub></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.504</td> <td>0.660</td> <td>0.464</td> <td>0.355</td> <td>0.692</td> <td>0.630</td> <td>0.543</td> <td>0.084</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">MMLU-Pro<sub>test</sub></th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.307</td> <td>0.475</td> <td>0.199</td> <td>0.286</td> <td>0.441</td> <td>0.428</td> <td>0.298</td> <td>0.099</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">Multi-If</th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.373</td> <td>0.470</td> <td>0.443</td> <td>0.304</td> <td>0.687</td> <td>0.523</td> <td>0.464</td> <td>0.236</td> </tr> <tr> <th style="text-align: left; font-weight: normal;">IFEval</th> <td style="background-color: rgba(127, 127, 127, 0.08);">0.749</td> <td>0.725</td> <td>0.776</td> <td>0.543</td> <td>0.869</td> <td>0.734</td> <td>0.679</td> <td>0.501</td> </tr> </tbody> </table> <p><small><sup>†</sup> Averaged over RefCOCOval, RefCOCOtestA, RefCOCOtestB, RefCOCO+val, RefCOCO+testA, RefCOCO+testB, RefCOCOgval, RefCOCOg_test.</small></p> <p><small><sup>‡</sup> SmolVLM2.2B's GQA output was scored as 0.000 under VLMEvalKit's answer-extraction rules.</small></p>
Citation
@misc{cohere_north_micro_vision_instruct,
title = {{North Micro Vision}: A 2.4B Native-Resolution Vision-Language Model},
url = {https://huggingface.co/blog/CohereLabs/meet-north-micro-vision-instruct},
author = {{Team Cohere}},
month = {August},
year = {2026}
}Contact
For errors or questions about this model card, contact Cohere Labs.