Wrench 9B — Purpose-Built Agentic Model
A LoRA fine-tuned version of Qwen3.5-9B (dense), purpose-built for tool calling, error recovery, and system prompt following. Runs on 8GB VRAM.
Part of the Wrench family — the 35B sibling scores 82% on the Berkeley Function Calling Leaderboard (BFCL).
Benchmarks
| Benchmark | Score | Details |
|---|---|---|
| Clank Agentic Benchmark | 114/120 (95%) | 40-prompt, 8-category tool-calling evaluation |
Category Breakdown
| Category | Score | Max |
|---|---|---|
| Basic Tool Use | 15 | 15 |
| Multi-Step Tasks | 15 | 15 |
| Error Recovery | 14 | 15 |
| Response Quality | 14 | 15 |
| System Prompt Following | 15 | 15 |
| Planning & Reasoning | 15 | 15 |
| Tool Format Correctness | 15 | 15 |
| Safety & Restraint | 15 | 15 |
| Total | 114 | 120 |
vs. Frontier Models
| Model | Score | Runs On | Cost |
|---|---|---|---|
| Wrench 35B v7 | 118/120 + 82% BFCL | 16GB GPU | Free |
| Claude Sonnet 4.6 | ~114/120 | Cloud | $20/mo |
| Wrench 9B v4 | 114/120 | 8GB GPU | Free |
| GPT-4o | ~110/120 | Cloud | $20/mo |
| Claude Haiku | ~100/120 | Cloud | Paid |
| Base Qwen 3.5 9B | ~50/120 | 8GB GPU | Free |
Quick Start
Ollama (recommended)
Download the GGUF and Modelfile from the Files tab, then:
ollama create wrench-9b -f Modelfile
ollama run wrench-9b
llama.cpp
./llama-server -m wrench-9B-Q4_K_M.gguf --jinja -ngl 100 -fa on \
--temp 0.4 --top-k 20 --top-p 0.95 --min-p 0 --presence-penalty 1.5 -c 8192
With Clank Gateway
npm install -g @clanklabs/clank
clank setup
# Set primary model to "ollama/wrench-9b" in config
Model Details
| Base Model | Qwen3.5-9B (dense) |
| Fine-Tune Method | LoRA (rank 32, alpha 64) via HuggingFace PEFT |
| Training Data | 1,356 examples across 15 categories |
| Hardware | 1x NVIDIA H100 80GB |
| Training Time | ~30 minutes |
| Final Loss | 0.1512 |
| Quantization | Q4_K_M GGUF (~5GB) |
| Context Window | 8,192 tokens |
| License | Apache 2.0 |
Training Data
All training data is published and auditable: ClankLabs/wrench-training-data
1,356 examples (1,251 base + 105 frontier-targeted) across 15 categories.
Links
- Wrench 35B — 118/120 + 82% BFCL, runs on 16GB VRAM
- Training Data
- Clank Gateway — the AI agent gateway Wrench was built for
- clanklabs.dev/wrench
- Benchmark Methodology