evomerge-pareto-gguf

evomerge-pareto-gguf

Pareto-optimal compressed LLMs across size constraints, in GGUF format. Ready to use with Ollama or llama.cpp.

Benchmark Results

• GSM8K: math reasoning (200 problems)
• IFEval: instruction following (94 problems)
• MMLU: general knowledge (570 problems, 57 subjects)

Which model to choose

Usage

Ollama (recommended)

One-command install with optimized settings:

# Download Modelfile then create
curl -L https://huggingface.co/tellerlin/evomerge-pareto-gguf/resolve/main/Modelfile-qwen3-8b-v2 -o Modelfile
ollama create evomerge-qwen3-v2 -f Modelfile
ollama run evomerge-qwen3-v2

Or run directly (default settings):

ollama run hf.co/tellerlin/evomerge-pareto-gguf:qwen3-8b-lora-v2-Q3_K_M

⚠️ If you use a custom Modelfile: Do not include a TEMPLATE line. Adding TEMPLATE {{ .Prompt }} will override Qwen3's built-in chat template and break tool calling, thinking mode, and system prompts. Only set PARAMETER and SYSTEM — Ollama will automatically use the correct template from the GGUF file.

llama.cpp

llama-cli -m qwen3-8b-lora-v2-Q3_K_M.gguf -p "Your prompt here" -n 512 -c 32768

Python (llama-cpp-python)

from llama_cpp import Llama
llm = Llama("qwen3-8b-lora-v2-Q3_K_M.gguf", n_ctx=32768)
out = llm("Your prompt", max_tokens=512)
print(out["choices"][0]["text"])

Base Models

• Qwen/Qwen3-8B-Instruct
• Qwen/Qwen3-1.7B
• Qwen/Qwen2.5-1.5B-Instruct