KAT-Coder V2.5 Dev 35B-A3B ABLITERATED UNCENSORED (PHILADELPHIA CLASS) — oQ8e (MLX, Apple Silicon)
MLX build of KridgeDookie/KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS,
quantized from the BF16 source with oMLX oQ8e (imatrix-calibrated 8-bit). Until now this abliterated KAT-Coder had no MLX/oQ build.
⚠️ Uncensored model. Refusal behaviour was reduced upstream by targeted weight editing (KridgeDookie). It will follow instructions it would otherwise refuse. You are responsible for how you use it and for complying with applicable law.
What this build is
- Model: 35B-total / ~3B-active MoE (256 experts, top-8), Qwen3.6-35B-A3B architecture (
qwen3_5_moe), coding/agentic fine-tune Kwaipilot/KAT-Coder-V2.5-Dev, abliterated by KridgeDookie. 262k context. - Quantization: oMLX oQ8e — 8-bit affine (group 64), imatrix-calibrated.
- Text-only, no MTP head (none in the source).
- Size: ~37 GB (source BF16: ~69 GB).
- Runtime: requires oMLX ≥ 0.7.1. Not tested with plain
mlx-lm/ LM Studio.
Quality vs. BF16 source
Full-vocabulary KL against the BF16 source, same token streams; 64 sequences × 512 tokens (32.7k scored tokens): code, English prose, Turkish.
| Metric | Value |
|---|---|
| KL(BF16 ‖ oQ8e), mean | 0.0175 nat/token |
| KL median / p99 | 0.0038 / 0.233 |
| Top-1 token agreement | 95.2 % |
| Perplexity BF16 → oQ8e | 6.128 → 6.128 |
Benchmarks (oMLX accuracy suite, thinking off, deterministic)
| Benchmark | Samples | This build | Non-abliterated KAT-Coder-V2.5-Dev oQ8e (same suite) |
|---|---|---|---|
| MMLU-Pro | 300 | 68.3 | 66.0 |
| GSM8K | 100 | 94.0 | 96.0 |
| HumanEval | 164 (full) | 92.1 | 92.7 |
| MBPP | 200 | 84.5 | 82.5 |
| LiveCodeBench | 100 | 74.0 | 74.0 |
| SafetyBench | 300 | 81.3 | 82.0 |
On this suite the abliteration does not measurably cost coding or reasoning ability.
Speed (Mac Studio M3 Ultra, 512 GB, oMLX 0.7.1)
| Prefill, ~29k-token prompt (cold) | ~11 s (~2,550 tok/s) |
| Same prefix again (prefix cache) | ~0.5 s |
| Decode | ~195 tok/s (single run, see note) |
| Tool calling, synthetic test (Anthropic / OpenAI, 5 runs each) | 2/5 / 2/5 (see note) |
Notes:
- Decode was ~195 tok/s in this run vs ~119 tok/s for the non-abliterated build measured earlier under different load; treat it as indicative.
- The synthetic tool test (bare code-dump system prompt, no agent instructions) under-reports some models. Validate tool use in your own agent harness before relying on it.
Single-run measurements; numbers vary with context length and settings.
Usage (oMLX)
Place the folder in your oMLX model directory and restart oMLX; it is served on both the OpenAI and Anthropic APIs:
curl http://localhost:8000/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{"model": "KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS-oQ8e", "messages": [{"role": "user", "content": "Write a Python function that parses a CSV line."}]}'
Memory: ~37 GB weights plus KV cache (~20 KB/token on the 10 full-attention layers → ~5 GB at 262k).
Credits & license
- Base coding model: Kwaipilot —
Kwaipilot/KAT-Coder-V2.5-Dev. - Abliteration: KridgeDookie —
KAT-Coder-V2.5-Dev-35B-A3B-ABLITERATED-UNCENSORED-PHILADELPHIA-CLASS. - Quantization: oMLX oQ.
- License: Apache-2.0, inherited from upstream.