---

license: apache-2.0

language:

• ru
• en

base_model: wayzer2321/Luna-3.5-9B-merged-private

base_model_relation: quantized

pipeline_tag: text-generation

tags:

• gguf
• qwen3.5
• luna
• q8_0
• mtp

---

Luna v2 — Qwen3.5 9B Q8_0 with MTP

Public GGUF release dated 2026-10-05.

Luna is Wayzer's conversational character: informal slang, sarcasm,

surreal humor, emotional shifts, and creator lore.

Release identity

• Checkpoint: Luna v2.
• Base: unsloth/Qwen3.5-9B.
• Quantization: Q8_0.
• File: luna-3.5-9b-Q8_0-MTP.gguf.
• File size: approximately 9.11 GiB.
• MTP tensors: preserved.
• Source: merged safetensors.

This uses a different fine-tuned checkpoint from the older public Luna 9B Q4.

Comparing those files therefore compares both checkpoint and quantization.

Ollama: local import

Download the GGUF into the current directory and create Modelfile:

FROM ./luna-3.5-9b-Q8_0-MTP.gguf
PARAMETER temperature 0.75
PARAMETER top_p 0.92
PARAMETER repeat_penalty 1.08
PARAMETER num_predict 140
PARAMETER num_ctx 2048

Then run:

ollama create luna-v2 -f Modelfile
ollama run luna-v2

These sampling settings are a starting point, not the training parameters.

MTP and memory

Preserved MTP weights require a compatible inference implementation to be used.

An ordinary model import does not establish that MTP acceleration is enabled.

Runtime memory must also cover caches and backend buffers in addition to

the GGUF file. Speed depends on hardware, offloading, context, and backend.

Provenance

Converted from the merged Luna v2 checkpoint using llama.cpp conversion.