KAT-Coder-V2.5-Dev-OrnithVision

A multimodal code model combining KAT-Coder-V2.5-Dev's language backbone with Ornith-1.5-35B-A3B's vision tower.

Background

The original Kwaipilot/KAT-Coder-V2.5-Dev does not ship with a vision tower or MTP layers — it is a pure language model. This model adds vision capability by transplanting the vision tower from Ornith-1.5-35B-A3B into the KAT language backbone.

Credits

This model is a fusion of two existing projects:

• Language backbone: KAT-Coder-V2.5-Dev by Kwaipilot
• Vision tower: Ornith-1.5-35B-A3B by ornith-ai
• Quantization: AWQ INT4 variant by cyankiwi

What changed

Component	Source
Language model (35B MoE, 3B active)	KAT-Coder-V2.5-Dev (unchanged)
Visual tower (27-layer ViT)	Ornith-1.5-35B-A3B (transplanted)
Quantization	AWQ INT4 (group size 32)
Total parameters	~35B
Active parameters per token	~3B

Technical details

• Architecture: Qwen3_5MoeForConditionalGeneration
• Vision encoder: Qwen3_VisionTransformer, 27 blocks, hidden_size=1152, 16 attention heads, patch_size=16
• Projector: visual.merger — 2 linear layers + 1 norm, output hidden_size=2048
• Token id: image token = 248056
• Dtype: float16
• Context length: up to 128K (with appropriate KV cache settings)

Usage

from transformers import AutoModelForCausalLM, AutoProcessor

model = AutoModelForCausalLM.from_pretrained(
    "KennyVo/KAT-Coder-V2.5-Dev-OrnithVision",
    torch_dtype="auto",
    device_map="auto"
)
processor = AutoProcessor.from_pretrained("KennyVo/KAT-Coder-V2.5-Dev-OrnithVision")

# Example: image + text
messages = [{"role": "user", "content": [
    {"type": "image"},
    {"type": "text", "text": "Describe this image."}
]}]
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=512)

# vLLM serving
vllm serve KennyVo/KAT-Coder-V2.5-Dev-OrnithVision --dtype float16

Method

1. Loaded visual weights (model.visual.*) from both cyankiwi/KAT-Coder-V2.5-Dev-AWQ-INT4 and cyankiwi/Ornith-1.5-35B-A3B-AWQ-INT4 — 333 tensors, all shapes match
2. Verified all tensor names and shapes are identical between the two models
3. Replaced all 333 visual tensors in KAT with Ornith's weights
4. Preserved all 123,493 non-visual tensors (language model, MoE experts, MTP) unchanged
5. Saved as 5 safetensors shards (~23 GB total)

Benchmarks

Text and reasoning benchmarks

Dataset	Subset	Num	Accuracy, KAT-Coder-V2.5-Dev-OrnithVision	Accuracy, Tech2wild/Nex-N2.5-mini-Int4-Int8Mix	cyankiwi/Ornith-1.5-35B-A3B-AWQ-INT4
GSM8K	main	1319	95.8%	96.6%	96.7%
ARC	ARC-Easy	2376	99.1%	98.9%
ARC	ARC-Challenge	1172	96.9%	98%
ARC	OVERALL	3548	98.4%	98.6%
MMLU	college_computer_science (STEM)	100	97%
MMLU	computer_security (STEM)	100	90%
MMLU	machine_learning (STEM)	112	91.1%
MMLU	OVERALL	312	92.6%

Vision benchmarks

Dataset MMMU with subset	Num	Accuracy, KAT-Coder-V2.5-Dev-OrnithVision	Accuracy, Tech2wild/Nex-N2.5-mini-Int4-Int8Mix
Accounting	30	73.3%	70%
Agriculture	30	66.7%	56.7%
Architecture_and_Engineering	30	50%	63.3%
Art	30	73.3%	73.3%
Art_Theory	30	86.7%	90%
Basic_Medical_Science	30	70%	73.3%
Biology	30	83.3%	76.7%
Chemistry	30	66.7%	56.7%
Clinical_Medicine	30	80%	80%
Computer_Science	30	60%	66.7%
Design	30	80%	80%
Diagnostics_and_Laboratory_Medicine	30	50%	46.7%
Economics	30	83.3%	83.3%
Electronics	30	40%	46.7%
Energy_and_Power	30	60%	76.7%
Finance	30	73.3%	63.3%
Geography	30	53.3%	66.7%
History	30	86.7%	76.7%
Literature	30	90%	93.3%
Manage	30	50%	53.3%
Marketing	30	83.3%	90%
Materials	30	56.7%	63.3%
Math	30	73.3%	70%
Mechanical_Engineering	30	53.3%	53.3%
Music	30	30%	23.3%
Pharmacy	30	93.3%	86.7%
Physics	30	96.7%	100%
Psychology	30	90%	83.3%
Public_Health	30	96.7%	93.3%
Sociology	30	76.7%	73.3%
OVERALL	900	70.9%	71%

Reproduction

Want to reproduce this model? See REPRODUCE.md for step-by-step instructions, or grab the transplant script transplant_ornith_vision.py and run it against the two source models.

Note: The entire combining study, Python source code, forensic analysis, and this model card were generated by KAT-Coder-V2.5-Dev itself — a model doing its own repair.

License

Apache 2.0 (inherited from base models)