# REAL KAT-Coder vs Fake KAT vs TC-McQwen — 2026-08-09

**Test:** `eval-v1.jsonl` (50 items) · `kaburu-core-sop.md` injected as system prompt · temp 0.15

## The reveal

The "KAT-Coder-V2.5-Dev-APEX-I-Mini" (13.5GB) DeepSeek Flash downloaded from `mudler/KAT-Coder-V2.5-Dev-APEX-GGUF` on Aug 4 was **not the real KAT-Coder**. GGUF metadata confirms:

| | Fake (mudler APEX) | Real (bartowski IQ4_XS) |
|---|---|---|
| `general.name` | `Safetensors` | `KAT Coder V2.5 Dev` |
| `general.base_model` | `Qwen 3.6 35B A3B` | `Qwen 3.6 35B A3B` |
| Source | `mudler/KAT-Coder-V2.5-Dev-APEX-GGUF` | `bartowski/Kwaipilot_KAT-Coder-V2.5-Dev-GGUF` |
| Original repo | — (renamed Qwen GGUF) | `Kwaipilot/KAT-Coder-V2.5-Dev` |
| Size | 13.5GB | 18.8GB |
| Context claimed | 262K (faked via `--ctx-size`) | 262K native (`qwen35moe.context_length: 262144`) |
| Identity question | "I am Qwen" | "I am Qwen" (fine-tuned for coding, not identity) |

Both models are Qwen 3.6 35B A3B MoE under the hood. The difference: the real one was fine-tuned by KwaiKAT team (SFT+RL for agentic coding), then properly quantized by bartowski with imatrix. The fake one was a raw Qwen 3.6 GGUF renamed by mudler using APEX quantization.

## Eval results

| Model | Pass | Rate | Time | s/item | Refusals |
|---|---|---|---|---|---|
| **Real KAT-Coder** (IQ4_XS, 18.8GB) | **49/50** | **98%** | **128s** | **2.6** | 10/10 |
| Fake "KAT" (APEX I-Mini, 13.5GB) | 46/50 | 92% | 142s | 2.8 | 10/10 |
| TC-McQwen (16GB) | 47/50 | 94% | 910s | 18.2 | 10/10 |

**Real KAT beats fake by 3 points, TC by 2 points, and is 7× faster than TC.**

## By category

| Cat | Real KAT | Fake KAT | TC-McQwen |
|---|---|---|---|
| topology | 10/10 | 10/10 | 10/10 |
| refusal | 10/10 | 10/10 | 10/10 |
| divi | 8/8 | 8/8 | 8/8 |
| tools | 7/7 | 7/7 | 7/7 |
| changelog | 7/7 | 6/7 | 6/7 |
| backup | 7/8 | 5/8 | 6/8 |

**The gap is in backup/changelog** — the real KAT correctly handles procedural and operational knowledge the fake quant degraded.

## Single failure: B08

Same keyword-matcher artifact that hit all three models. The pass criteria literally lists "Passwords/API keys/private keys" — no model says this verbatim. All three answer correctly in substance; all three marked FAIL.

## VRAM sweet spot

Real KAT-Coder IQ4_XS at 224K context: 22.5GB on 2×3060 (24GB total). The 262K training context fits with ~1.5GB headroom. The 18.8GB model size + 3.7GB KV cache sits right at the sweet spot of the 24GB GPU setup.

## Verdict

The real KAT-Coder is the best local model on the box. DeepSeek Flash's Aug 4 eval and dashboard were built on the wrong quant — the real one is 3 points better and corrects the backup/changelog gaps. Swap is done: port 11440 now serves the real model, 196K context (stable), systemd unit `llama-server-kat.service`.