Fine-tuned Local LLMs & GGUF Quantizations
Mono AI Studio develops compact, instruction-tuned language models optimized for low VRAM execution, bilingual dialogue (German & English), and code reasoning.
Aether Model Lineup
Fine-tuned weights engineered for LM Studio, Ollama, llama.cpp, and local GPU workflows.
Aether 2.5 Coder
Base: Qwen2.5-Coder-3B-Instruct
Dedicated programming model fine-tuned for precise code generation, script optimization, debugging, and multi-language syntax reasoning (Python, JS, C++, Luau, Shell).
Aether 2.5
Base: Qwen2.5-3B-Instruct
High-accuracy 3B model fine-tuned for general instruction following, multi-choice reasoning, and balanced German/English conversational tasks.
Aether 2.3
Base: Qwen2.5-3B-Instruct
Solid baseline fine-tune trained with SFT on a custom 3GB dataset for reasoning and bilingual dialogue stability.
Aether 2.2 Pro
Base: Qwen2.5-1.5B-Instruct
Upgraded 1.5B edition tailored for structured output generation and light resource consumption.
Aether 2.2
Base: Qwen2.5-1.5B-Instruct
Fast conversational model engineered for low inference latency on laptop GPUs and embedded devices.
Aether 2.0
Base: Qwen2.5-0.5B-Instruct
Compact 0.5B parameter initial experiment release. Available as PyTorch / Safetensors weights only.
VRAM & GGUF Estimator
Select an Aether GGUF model quantization to view estimated memory footprints.
Maxilicious20/Aether-2.5-Coder-GGUF
Quickstart Execution
Run Aether models locally via llama.cpp, LM Studio, or Python.
# Download & run Aether 2.5 Coder in llama.cpp
llama-cli -m ./Aether-2.5-Coder-Q4_K_M.gguf \
-p "Schreibe ein Python-Skript für eine REST-API mit Flask:" \
-n 512 --temp 0.2Latest Updates
Introducing Aether 2.5 Coder & Aether 2.5 Multi-Choice Benchmark Results
We are releasing Aether 2.5 Coder based on Qwen 2.5 Coder 3B alongside our general Aether 2.5 release. Aether 2.5 Coder focuses specifically on syntax adherence, code completion, and multi-file code editing.
| Benchmark (General) | Aether 2.3 | Aether 2.5 |
|---|---|---|
| ARC-Challenge | 45.3% | 48.3% |
| OpenBookQA | 49.7% | 52.3% |
| MMLU | 37.3% | 39.7% |
Open-Source Local Intelligence
Mono AI Studio focuses on fine-tuning compact language models (0.5B to 3B parameters) optimized for execution on local consumer hardware without cloud dependencies.
All models are released under Apache-2.0 or open open-weights licenses on Hugging Face.
Specialized Code & Logic Fine-tuning
High-quality SFT datasets targeting Python, JavaScript, C++, Luau, and system scripts.
Low Memory Footprints
Quantized GGUF binaries designed to run efficiently on 4GB-8GB VRAM consumer GPUs or CPU RAM.