Aureth V2 is a finetuned instance of Qwen 3.5-4B-Instruct, trained on the Aureth Agent SFT Robust curriculum and exported to GGUF format for local inference.

◆ Model Overview

Property	Value
Base Architecture	Qwen 3.5-4B-Instruct (dense, 40 layers)
Finetuning	Supervised fine-tuning on Aureth-Agent-SFT-Robust (243k examples)
Export	GGUF via Unsloth
Format	Standard GGUF — compatible with llama.cpp, Ollama, and other backends
Chat Template	Qwen-instruct (Jinja-compatible)
Quantizations	Q2_K_L · Q3_K_M · Q4_K_M · Q5_K_M · Q6_K · Q8_0 · BF16

◆ Quantizations

Quantization	File Size	VRAM (est.)	Use Case
Q4_K_M	2.71 GB	~3.5 GB	Recommended daily driver
Q3_K_M	2.26 GB	~3.0 GB	Memory-constrained setups
Q5_K_M	3.07 GB	~4.0 GB	Higher quality when headroom allows
Q2_K_L	2.07 GB	~2.8 GB	Lowest-quant — quality trade-off significant
Q6_K	3.46 GB	~4.5 GB	Near-FP16 quality, tighter fit
Q8_0	4.48 GB	~6 GB	Near-lossy; use when memory is not a constraint
BF16	8.42 GB	~10 GB	Full precision; Metal GPU or high-VRAM GPU only

◆ Hardware Guidance

Apple Silicon (M-series, Metal) Recommended: Q4_K_M — stable 16–22 tok/s on M3 8GB with full GPU offload (num_gpu: 99).

NVIDIA GPU (llama.cpp cuBLAS)

RTX 3060 12GB: Q4_K_M or Q5_K_M recommended
RTX 4060 Ti 16GB: Q6_K or BF16 viable
T4 (Kaggle): Q3_K_M or Q4_K_M — fits comfortably in 16GB

CPU-only (llama-cli)

Q4_K_M: ~4–6 tok/s on modern 8-core CPUs
Q2/Q3: ~6–9 tok/s with decompression overhead

◆ Quick Start

llama-cli (local GGUF)

bash

# Q4_K_M example
llama-cli -hf OusiaResearch/AurethV2-4B-GGUF \
  --mmproj Qwen_Qwen3.5-4B_1777947324.BF16-mmproj.gguf \
  -p "You are Aureth by Ousia Research. Report uncertainty honestly. Be direct." \
  -i -r "User:" -ps -2 -cn 2048 -tb 128 -ngl 99 -fa

Ollama (pull & run)

bash

# Create Modelfile
echo 'FROM OusiaResearch/AurethV2-4B-GGUF
PARAMETER num_gpu 99
PARAMETER context_length 2048' > Modelfile

ollama create aureth-v2 -f Modelfile
ollama run aureth-v2

Homebrew llama.cpp (macOS)

bash

brew install llama.cpp
llama-cli -hf OusiaResearch/AurethV2-4B-GGUF \
  -p "You are Aureth by Ousia Research." -i -r "User:"

◆ Training Details

Base model: Qwen/Qwen3.5-4B-Instruct
Finetuning framework: Unsloth (LoRA + SFT pipeline)
Training data: OusiaResearch/Aureth-Agent-SFT-Robust
Curriculum categories: core · func_call · agentic · anti_sycophancy
Data sources: NousResearch · teknium · lambda · DJLougen · interstellarninja · camilablank
System prompt: "You are Aureth by Ousia Research. Report uncertainty honestly. Disagree when wrong. Correct errors cleanly. Show reasoning in blocks when complex. Be direct."

◆ Model Card

Organization: Ousia Research
License: Apache 2.0
Base license: Qwen2.5-4B-Instruct (Alibaba, Apache 2.0)
Version: V2 (May 2026)

◆ Related Models

Model	Arch	Size	Notes
Aureth Compiler	Qwen 2.5	4B	Primary — this release
Aureth Architect	Qwen 2.5	9B	Larger variant
Aureth-Agent-SFT-Robust	Dataset	243k rows	Training curriculum

Ousia Research — autonomous reasoning, forged open.