---

tags:

• gguf
• quantized
• mtp
• qwen3.5

base_model: Lolol857/merged-9b-linear-60-25-15-mtp

license: apache-2.0

language:

• en

---

Merged 9B Linear 60/25/15 + MTP (GGUF)

GGUF quantized versions of Lolol857/merged-9b-linear-60-25-15-mtp.

Original merge composition

• 60% empero-ai/Qwythos-9B-v2 (reasoning + chat)
• 25% kai-os/Carnice-9b (agentic)
• 15% allenai/tmax-9B (agentic)

All three are Qwen3.5-9B fine-tunes, linearly merged. The MTP head was copied

from empero-ai/Qwythos-9B-v2 (the only source that had MTP tensors).

Details

Slightly wrong in the head, but usable. engineered for better agentic capability.

No benchmarks yet.

Available quantizations

| File | Format | Size | Notes |

|------|--------|------|-------|

| merged-9b-mtp.Q4_K_M.gguf | Q4_K_M | ~5.5 GB | Most popular, good balance |

| merged-9b-mtp.Q5_K_M.gguf | Q5_K_M | ~6.5 GB | Higher quality |

| merged-9b-mtp.Q8_0.gguf | Q8_0 | ~9.5 GB | Near-lossless |

Usage with llama.cpp

llama-server -m merged-9b-mtp.Q4_K_M.gguf -c 8192 --port 8080

Usage with Ollama (can't think right)

Create a Modelfile:

FROM ./merged-9b-mtp.Q4_K_M.gguf

Then: ollama create merged-9b-mtp -f Modelfile

Disclaimer

This is an automated merge + MTP attachment. Quality depends on the

compatibility of source weight spaces. Evaluate on your task before deployment.