# MTP Model Downloads — unsloth repos

## How unsloth MTP repos work

The unsloth MTP repos (e.g. `unsloth/Qwen3.5-9B-MTP-GGUF`) contain the MTP-quantized models directly — the Q4_K_M file in the repo IS the MTP version. You don't need a separate "MTP" quant; the repo itself is the MTP variant.

## Downloaded models

### Qwen3.5-9B-MTP
- **Repo:** `unsloth/Qwen3.5-9B-MTP-GGUF`
- **File:** `Qwen3.5-9B-Q4_K_M.gguf` (5.5 GB)
- **Local path:** `/models/downloads/Qwen3.5-9B-MTP-Q4_K_M.gguf`
- **llama-swap key:** `qwen35-9b-mtp`
- **Aliases:** uncensored, qwen35-ud, qwen35-9b, Qwen3.5-9B-UD
- **Flags:** MTP enabled (`--spec-type mtp --spec-draft-n-max 2`), `--no-kv-offload`, `--temp 0.7 --top-p 0.8`
- **Benchmark (2080 Ti @ 200K):** 33.6 t/s gen, 252 t/s PE, 6.9GB VRAM

### Qwen3.6-35B-A3B-MTP
- **Repo:** `unsloth/Qwen3.6-35B-A3B-MTP-GGUF`
- **File:** `Qwen3.6-35B-A3B-UD-Q4_K_M.gguf` (21.1 GB)
- **Local path:** `/models/downloads/Qwen3.6-35B-A3B-MTP-Q4_K_M.gguf`
- **llama-swap key:** `qwen36-35b-mtp`
- **Aliases:** qwen-ud, qwen36-ud, qwen36-35b, Qwen3.6-35B-UD
- **Flags:** MTP enabled, `--n-cpu-moe 128`, `--temp 0.2 --top-p 0.95`, batch=256/256
- **Benchmark (2080 Ti @ 200K):** 33.8 t/s gen, 74 t/s PE, 8.9GB VRAM

### Heretic-35B (native MTP)
- **Local path:** `/models/downloads/Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved.Q4_K_M.gguf` (21 GB)
- **llama-swap key:** `heretic-35b` (daily driver)
- **Flags:** MTP enabled, `--n-cpu-moe 128`, `--threads 32`, batch=256/256
- **Benchmark (2080 Ti @ 250K):** 36.6 t/s gen, 90 t/s PE, 4.2GB VRAM

## Key difference from heretic variant

The `heretic-35b` model uses the **heretic-native-MTP-preserved** quant (`Qwen3.6-35B-A3B-uncensored-heretic-Native-MTP-Preserved.Q4_K_M.gguf`), which is a different quantization variant from the unsloth MTP. Both have MTP but differ in quantization scheme and training data. The unsloth MTP is the "clean" unsloth variant; heretic is the custom-tuned variant.

## Download command

```bash
# Qwen3.5-9B-MTP
curl -sL -o /models/downloads/Qwen3.5-9B-MTP-Q4_K_M.gguf \
  "https://huggingface.co/unsloth/Qwen3.5-9B-MTP-GGUF/resolve/main/Qwen3.5-9B-Q4_K_M.gguf"

# Qwen3.6-35B-A3B-MTP
curl -sL -o /models/downloads/Qwen3.6-35B-A3B-MTP-Q4_K_M.gguf \
  "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF/resolve/main/Qwen3.6-35B-A3B-UD-Q4_K_M.gguf"
```

## Benchmark results (2080 Ti, May 2026)

| Model | Context | Gen t/s | PE t/s | VRAM | Load time |
|---|---|---|---|---|---|
| heretic-35b | 250K | 36.6 | 90 | 4.2GB | 64s |
| qwen36-35b-mtp | 200K | 33.8 | 74 | 8.9GB | 44s |
| qwen36-35b-mtp | 250K | 34.7 | 64 | 9.3GB | 35s |
| qwen35-9b-mtp | 200K | 33.6 | 252 | 6.9GB | 6s |