Open your business account with Statrys 👉 https://bit.ly/StatrysWebsite0:00 Intro1:48 The Switch Nobody Talks About3:53 The Price War You're Losing5:28 The ...
Anthropic banned me for no reason last week, was merely using it for some SaaS project. Appealed, nothing. No reason given other than usage policy blabla.
Google Gemini, unusable, timed out 9 out of 10 times.
Very expensive, and unsustainable in my opinion. All the big boys lack capacity and I doubt they will be able to get that sorted.
Now running Qwen3.6:27B, opensourced by Ali Baba, locally. Works really well actually.
It’s honestly shocking how good Qwen 3.6:27b is, it actually outperforms Qwen 3.5:397b which itself was released only in February. I’m convinced that local models are the future. In a year or two, we’ll probably get to the point where local models are as good as Claude is today, and at that point it kind of doesn’t even really matter if frontier keeps getting better. It’s going to be good enough for vast majority of use cases. On top of that, you can already do a lot with tooling to make the model perform better. This paper is a great example. And I think this is very much an underexplored area. Current agentic harnesses are very primitive in nature, they just give the model some tools to play with, but do little in a way of guidance. ATLAS is a really interesting project in that’s attempting to make a smarter harness, and their results are pretty impressive. If you’re already running Qwen locally, I recommend checking it out.
Anthropic banned me for no reason last week, was merely using it for some SaaS project. Appealed, nothing. No reason given other than usage policy blabla.
Google Gemini, unusable, timed out 9 out of 10 times.
Very expensive, and unsustainable in my opinion. All the big boys lack capacity and I doubt they will be able to get that sorted.
Now running Qwen3.6:27B, opensourced by Ali Baba, locally. Works really well actually.
It’s honestly shocking how good Qwen 3.6:27b is, it actually outperforms Qwen 3.5:397b which itself was released only in February. I’m convinced that local models are the future. In a year or two, we’ll probably get to the point where local models are as good as Claude is today, and at that point it kind of doesn’t even really matter if frontier keeps getting better. It’s going to be good enough for vast majority of use cases. On top of that, you can already do a lot with tooling to make the model perform better. This paper is a great example. And I think this is very much an underexplored area. Current agentic harnesses are very primitive in nature, they just give the model some tools to play with, but do little in a way of guidance. ATLAS is a really interesting project in that’s attempting to make a smarter harness, and their results are pretty impressive. If you’re already running Qwen locally, I recommend checking it out.
Personallly, I think qwen3.6 is already pretty close to Claude.
Did not know about ATLAS, will definitely try that out.
It’s really getting there, it’s the first local model that really does feel like frontier quality.