Google's Gemma 4 31B FP8 matches Anthropic's Sonnet 4.6 on complex tasks
A 31B quantized model rivals a much larger proprietary counterpart on graph queries and code.
Deep Dive
Cypher graph traversal for Neo4j, entity extraction from text chunks (web, graph, vectors), agentic tool calling (Skills selection for Pi), code writing in Python, and synthesis/summarization of multi-vector retrieval—all with Gemma/Qwen in FP8. This brought the author joy.
Key Points
- Gemma 4 31B FP8 matched Claude Sonnet 4.6 medium on Cypher graph traversal, entity extraction, and agentic tool calling.
- The custom benchmark also covered Python code writing and multi-vector retrieval summarization.
- FP8 quantization allows a 31B model to run efficiently while retaining near-proprietary-level performance.
Why It Matters
This shows open-weight 31B models can rival proprietary 100B+ models, enabling cost-effective local AI for complex agentic workflows.