I Tried CodeBuddy’s Free Hy3 for One Day — It’s Behind the Domestic Models
TRAE killed its free tier, so I spent a day on CodeBuddy’s free Hy3 model doing light doc tasks. The honest verdict: Hy3 lags behind the domestic models I’ve used, and long tasks cost too much rework. Here’s the day’s record.
Background / Problem
TRAE recently changed its pricing and removed the free tier. Since I knew CodeBuddy’s Hy3 was free, I spent today putting it through some light document-work tasks — partly to see if the free tier is good enough, and partly to evaluate Hy3 itself.
What happened in the day
Today I tried CodeBuddy + Hy3 on some simple document-flow tasks I’d described earlier.
The feeling: Hy3 is still behind the domestic models I’ve used.
My usual habit is long tasks — I write the requirements clearly and hand them over in one pass. This time the output had too many problems: it didn’t fully understand the original content, and it didn’t fully understand my needs. I spent a lot of time reworking and fixing.
Hy3 has a long way to go, and right now it’s visibly behind.
Plan for tomorrow
Tomorrow I’ll try some simple coding tasks and see how it does.
But one thing is already clear: I won’t hand it long tasks. Otherwise I can’t control the quality myself. Long tasks stay on the models with steadier comprehension.
Context: my tool rotation
A few months ago I tried most domestic tools and free tiers. CodeBuddy itself is pretty good — Tencent’s UX is solid, and it also plugs in third-party models. Last month I subscribed to TRAE for a deep month; this month it’s Qoder; next month I plan to subscribe to CodeBuddy for a proper month-long test. The domestic models and tools are genuinely competing and growing.
Mistakes, tradeoffs, alternatives
- The trap: a “free” model stops being free once you’re fixing its mistakes for an hour. Rework cost is the real price.
- Why not just use it for everything: long-task comprehension is where stronger (often paid) models earn their keep.
- When I would NOT recommend Hy3 yet: any multi-step task where you can’t easily verify and control the output.
Conclusion & further reading
- Hy3 runs simple doc tasks but lags behind the domestic models I’ve used.
- Long-task rework is high; I won’t trust it with long tasks yet.
- Tomorrow: simple coding test. Next month: a deeper CodeBuddy subscription trial.
Did you hit a gotcha I missed while evaluating free AI coding models? Drop it in the comments and I’ll reply to every one.
Note: This very post was also generated by CodeBuddy + Hy3 from my rough thoughts. Honestly, it’s missing a little something — but I don’t have time to polish it, so here it is.
This is the only time an AI failed to read the AI-skill correctly and generated the blog in the wrong directory. After I told it to fix this, instead of simply moving the files, it oddly regenerated all twelve platform versions from scratch. Previously, for my blogs I’d just grab whatever was at hand — ClaudeCode+DeepSeek, TRAE+Seed, TRAE+Auto, Qoder+Auto…