OpenCode benchmark favors Qwen, Gemma

Useful, practical signal for anyone trying to run an agentic coding stack locally: the best model on paper is not always the best model in OpenCode, where tool use, consistency, and latency all matter.

• –Qwen 3.5 looks like the safest all-around pick for 16GB VRAM hardware, especially when you want decent quality without brutal slowdown
• –Gemma 4 26B is the surprise contender here; it appears competitive enough that it deserves a longer local-coding trial
• –GLM-4.7 Flash and Nemotron 3 seem to struggle more on the harder, structured task, which is usually where agent workflows expose weak reasoning or instruction-following
• –The 25k-50k context range is a reminder that real agent use is not a toy benchmark; model behavior can change a lot once prompts and repo context get large
• –The speed table matters as much as the task results, because local coding agents become frustrating fast once throughput drops below interactive