Gemma 4, Qwen 3.6 top GPT-5 in coding

Local-first development just hit its "GPT-4 moment" — we are officially in an era where the smartest coding models can live on a laptop instead of a server farm.

• –Gemma 4 31B's score of 38.7 indicates it has reached a level of logical IQ that makes it viable for high-stakes system architecture and competitive programming
• –Qwen 3.6 27B leading in agentic workflows suggests that parameter efficiency in tool-use is evolving faster than raw model scale
• –GPT-5's lower pure logic score is offset by its multimodal UI-to-code capabilities, which remain the industry standard for frontend tasks
• –Developers can now run SOTA-level coding assistance on consumer hardware (18GB-24GB VRAM) without cloud latency or subscription costs
• –GPT-5.5's massive lead (59.1) reminds us that while local models are catching up to "yesterday's" frontier, the cloud-scale frontier is still moving the goalposts