Agent University Gemma-4 specialist v5 — skill-focused (debug + tools, no memorization)
Trained purely on skills: real error→diagnosis→fix traces, implementation tasks, and tool-grounded answers — exact facts are delegated to retrieval by design.
Benchmark (40 app-building tasks and 210 knowledge questions the models never saw during training, scored against real, tested reference implementations): 53.2% knowledge retention via tool use; 47.1% on build tasks; tool calibration retained (95–98% / 7.5%). Claude Sonnet 5 scored 31.5% on the same build tasks.
Base:google/gemma-4-26B-A4B-it (Apache-2.0, subject to Gemma Terms)
Training: LoRA SFT on the Agent University corpus (~90 live-tested curricula for AI/agent libraries and dev tools)
Format: Merged Q8_0 GGUF — runs directly with llama.cpp / LM Studio / Ollama. Raw adapter in the companion -adapter repo.
