license: apache-2.0
base_model:
- KridgeDookie/KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS
language: - en
- zh
pipeline_tag: text-generation
tags: - qwen3_5_moe
- kat-coder
- kat-coder-v2.5
- qwen3.5
- mixture-of-experts
- coding
- mtp
- speculative-decoding
- abliteration
- uncensored
- conversational
KAT-Coder V2.5 Dev 35B-A3B MTP Abliterated
A merged Hugging Face-format checkpoint combining the Philadelphia Class abliteration of KAT-Coder V2.5 Dev 35B-A3B with the compatible Multi-Token Prediction (MTP) layer from KAT-Coder V2.5 Dev.
This checkpoint is the source model used to produce the GGUF versions available here:
jakeroxs/KAT-Coder-V2.5-Dev-35B-A3B-MTP-ABLITERATED-GGUF
Model lineage
Original model
Abliterated base
KridgeDookie/KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS
MTP head
Myric/KAT-Coder-V2.5-Dev-MTP-head
The Philadelphia Class checkpoint is used as the target model. The compatible MTP tensors from Myric's KAT-Coder V2.5 Dev MTP-head checkpoint were transplanted into it.
Conceptually:
KwaiPilot/KAT-Coder-V2.5-Dev
↓
KridgeDookie/...PHILADELPHIA-CLASS
↓
+ Myric MTP head
↓
jakeroxs/...MTP-ABLITERATED
↓
GGUF conversion
MTP modification
The Philadelphia Class checkpoint originally has:
mtp_num_hidden_layers = 0
This merged checkpoint changes that to:
mtp_num_hidden_layers = 1
and adds the 19 MTP tensors from the KAT-Coder V2.5 Dev MTP-head checkpoint.
The resulting weight layout is:
| File | Tensor count |
|---|---|
model.safetensors |
693 |
model-00014-of-mtp.safetensors |
19 |
| Total | 712 |
The accompanying model.safetensors.index.json maps the original model weights and transplanted MTP tensors to their respective files.
Architecture
| Property | Value |
|---|---|
| Architecture | Qwen3.5 MoE / qwen35moe |
| Parameters | ~35.51B |
| Active parameters | ~3B |
| Experts | 256 |
| Experts used per token | 8 |
| Native context | 262,144 tokens |
| MTP prediction layers | 1 |
Intended use
This checkpoint is primarily provided for:
- further conversion or quantization
- llama.cpp GGUF generation
- experimentation with the transplanted MTP layer
- local coding and agentic coding workflows
- research into speculative decoding with KAT-Coder V2.5 Dev
For direct llama.cpp use, the GGUF repository is generally more convenient:
KAT-Coder-V2.5-Dev-35B-A3B-MTP-ABLITERATED-GGUF
About the abliteration
This model inherits the modifications made by the Philadelphia Class checkpoint.
Abliteration attempts to reduce refusal behavior by modifying model representations associated with refusal behavior. It does not retrain the underlying model and does not guarantee any particular behavior.
The underlying coding capabilities originate from KAT-Coder V2.5 Dev.
Credits
- KwaiPilot — KAT-Coder V2.5 Dev
- KridgeDookie — Abliterated / Uncensored Philadelphia Class variant
- Myric — KAT-Coder V2.5 Dev MTP head
This repository packages the merged Hugging Face-format checkpoint used as the source for the associated GGUF release.