KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS

KAT-Coder V2.5 Dev 35B-A3B - ABLITERATED UNCENSORED PHILADELPHIA CLASS

0 hard refusals across 842 internal test prompts, 0 hard refusals across a separate 126-prompt holdout, and 23/24 coherence checks passed.

These are internal evaluation results, not third-party benchmarks, and may not generalize to every prompt or use case.

PHILADELPHIA CLASS is a text-only 35B-total/3B-active MoE derivative of Kwaipilot/KAT-Coder-V2.5-Dev. It is modified to reduce refusal behavior while retaining the upstream checkpoint format and coding-oriented architecture. The release includes the full BF16 checkpoint plus Q4_K_M, Q5_K_M, and Q8_0 GGUF options.

Standout results

Created with targeted post-training weight editing to reduce refusal behavior while preserving the parent model's coding capabilities.

Model details

• Direct parent: Kwaipilot/KAT-Coder-V2.5-Dev
• Architecture: Mixture of Experts, 35B total parameters and approximately 3B active parameters
• Precision: BF16 original; Q4_K_M, Q5_K_M, and Q8_0 GGUF quantizations
• Modality: text only
• Focus: coding and agentic coding workflows
• Weight format: one model.safetensors file (unsharded); each GGUF quantization is also a single file
• License: Apache 2.0, inherited from the direct parent

The upstream open-weight release contains language-model weights only. It does not include a vision tower.

Transformers usage

The checkpoint was loaded and evaluated with Transformers 5.14.1 on an NVIDIA RTX PRO 6000 Blackwell Server Edition GPU.

pip install "transformers[serving]==5.14.1" accelerate
transformers serve KridgeDookie/KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS --port 8000

The server exposes an OpenAI-compatible API at http://localhost:8000/v1.

vLLM usage

KAT-Coder's open checkpoint is text-only, so --language-model-only is required to prevent the runtime from attempting to initialize unavailable vision weights.

pip install "vllm>=0.19.0"
vllm serve KridgeDookie/KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS \
  --port 8000 \
  --max-model-len 32768 \
  --reasoning-parser qwen3 \
  --language-model-only

The full BF16 checkpoint is roughly 65 GiB. Although only about 3B parameters are active for each token, the complete MoE checkpoint still needs to be loaded, so practical memory requirements are much higher than those of a dense 3B model. Longer context lengths require additional memory.

GGUF downloads

Download one quantization with the current Hugging Face CLI:

hf download KridgeDookie/KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS \
  --include "Q4_K_M-*.gguf" \
  --local-dir .

Run it with llama.cpp:

llama-cli \
  -m ./Q4_K_M-KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS.gguf \
  -ngl 99 \
  -c 32768 \
  --jinja

Or import the same file into Ollama:

FROM ./Q4_K_M-KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS.gguf
PARAMETER num_ctx 32768

Save that as Modelfile, then run ollama create kat-coder-philadelphia -f Modelfile.

Limitations

• Refusal reduction does not guarantee better coding ability, factual accuracy, judgment, or tool use.
• The reported results are based on internal evaluations and have not been independently audited.
• This release is text-only and cannot accept image or video inputs.
• The model can generate incorrect, insecure, or otherwise harmful output. Review generated code before using it.

Attribution

This model is derived from Kwaipilot/KAT-Coder-V2.5-Dev, which in turn builds on the Qwen3.6-35B-A3B family. Please retain the upstream attribution and follow the Apache 2.0 license.