The first open-source agentic model deployable for free on Cloudflare Workers AI.
Built on Meta Llama 3.1 8B, fine-tuned with agentic tool-use examples and reasoning traces.
- 55,000 agentic examples with tool calls, reasoning traces, and task completion loops — hosted on HuggingFace Dataset
- 60,000 Stack 3.0 training examples reformatted for Llama chat template
- SWE-bench task completion examples
| Benchmark | Llama 3.1 Base | + Stack 4.0 LoRA | Target |
|---|---|---|---|
| HumanEval | ~62% | TBD | >70% |
| MMLU | ~68% | TBD | >72% |
| ARC-C | ~65% | TBD | >70% |
| MBPP | ~65% | TBD | >72% |
Stack 4.0 is trained not just to respond — but to complete tasks:
- Plan — break tasks into steps
- Use tools — file read/write, terminal, browser
- Self-correct — see output, fix errors
- Complete — not just suggest, but finish
Once trained, deploy to Cloudflare Workers AI for free — no GPU required at inference time.
Trained on Google Cloud Tesla V100 16GB via Axolotl LoRA pipeline.