Agent University Gemma-4 specialist v3 — recall drills + more tasks
Adds exact-identifier recall drills and 6× more implementation tasks to the v2 recipe.
Benchmark (40 app-building tasks and 210 knowledge questions the models never saw during training, scored against real, tested reference implementations): 53.3% on build tasks; best exact-fact retention of the SFT-only line (21.4%). For comparison, 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 supervised fine-tuning on the Agent University corpus — ~90 live-tested curricula for AI/agent libraries and dev tools (Supabase, MCP, Next.js, Cloudflare, Slack, and more)
Format: Merged Q8_0 GGUF — runs directly with llama.cpp / LM Studio / Ollama. The raw LoRA adapter is in the companion -adapter repo.
