gemma-4-26b-a4b-it-au-v4-adapter

Agent University Gemma-4 specialist v4 — tool-calling specialist

Trained to call a knowledge-base search tool (au_search) on its own: it looks facts up instead of guessing. Emergent calibration: calls the tool on 92–98% of knowledge questions and 0% of coding tasks.

Benchmark (40 app-building tasks and 210 knowledge questions the models never saw during training, scored against real, tested reference implementations): 49.6% knowledge coverage with self-directed tool use (≈2× Claude Sonnet 5 without tools); 52.0% on build tasks. 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: Fireworks fused_peft_3d_v1 layout — MoE expert LoRAs are fused 3D tensors. Standard PEFT loads the attention/MLP portion; for a full-fidelity merge use merge_adapter.py from the agent-university harness. A merged, ready-to-run Q8_0 GGUF is in the companion -gguf repo.