Update README.md

This model is derived from:

- cyankiwi/KAT-Coder-V2.5-Dev-AWQ-INT4 (Apache-2.0)
- cyankiwi/Ornith-1.5-35B-A3B-AWQ-INT4 (MIT)

The language model weights are inherited from KAT.
The visual tower (model.visual.*) is transplanted from Ornith.

Users must comply with the licenses of both upstream models.

# KAT-Coder-V2.5-Dev Vision Fix Experiment

I noticed that `cyankiwi/KAT-Coder-V2.5-Dev-AWQ-INT4` was producing poor image understanding results compared to `cyankiwi/Ornith-1.5-35B-A3B-AWQ-INT4`.

After inspecting the checkpoints:

- Both models contain 333 `model.visual.*` tensors
- Tensor names and shapes match 1:1
- KAT visual bias tensors are all zero
- Ornith visual bias tensors are non-zero

As an experiment, I replaced the entire `model.visual.*` subtree in KAT with the corresponding tensors from Ornith while keeping:

- KAT language model weights
- KAT tokenizer
- KAT config

unchanged.

Result:

✅ Model loads normally in vLLM

✅ Image understanding improved dramatically

✅ Coding capabilities remain unchanged (KAT language model preserved)

Statistics:

- Replaced tensors: 333
- Replaced parameters: 446,571,248 (~447M)
- Preserved non-visual tensors: 123,493

The transplant script is attached below for anyone interested in reproducing the experiment.

What the transplant changes
Component Sourcemodel.visual.* Ornith
model.language_model.* KAT
Tokenizer KAT
Config KAT

Statistics
Replaced visual tensors: 333
Replaced visual parameters: 446,571,248
Preserved non-visual tensors: 123,493
Language model weights unchanged

# Disclaimer

This is an unofficial experimental checkpoint.

Changes:
- Replaced all `model.visual.*` tensors in KAT-Coder-V2.5-Dev
with the corresponding tensors from Ornith-1.5-35B-A3B.
- Language model weights remain unchanged.

No retraining was performed.

README.md

@@ -1,3 +1,9 @@
 ---
 license: apache-2.0
----
+base_model:
+- Kwaipilot/KAT-Coder-V2.5-Dev
+- ornith-ai/Ornith-1.5-35B-A3B
+- cyankiwi/Ornith-1.5-35B-A3B-AWQ-INT4
+- cyankiwi/KAT-Coder-V2.5-Dev-AWQ-INT4
+pipeline_tag: image-text-to-text
+---