I’d agree that it’s a milestone in achievable local LLM coding performance on less than extreme hardware. I’ve been running a 3-bit quant on a 5070Ti and it’s the first local model that persuaded me to properly wire up a local agent into my dev flow.
However, it’s not enough to get me to drop my Claude (or other frontier model) sub and I don’t see it bursting the bubble*. It does certainly shift the landscape in terms of how much I’d be willing to pay and what I expect in terms of quality and performance for my sub though.
My experience will be different because “one-shot” tests are entirely meaningless to me for what I do. I have large projects with legacy codebases. Qwen 3.8 needs for more direction to deal with those than the like of Opus or Fable, especially when running on my constrained hardware.
*I’m really just using that term to reflect the post. I’m not someone who believes the dev side of LLMs is a bubble at this point, although datacenter hyperscaling and OpenAI in particular may well be.