Mimo v2.5 Pro의 다양한 면모
The many sides of Mimo v2.5 Pro
핵심 요약
Mimo v2.5 Pro를 직접 사용해 본 결과, 코딩 작업에서는 실망스러운 성능을 보였으나 페르소나 기반의 비평 능력은 뛰어남.
- 코딩 성능 — 간단한 HTML 생성 작업에서 반복적인 루프와 오류 발생
- 페르소나 비평 — 특정 역할을 부여했을 때 코드 개선 및 분석 능력 탁월
- 사용자 경험 — 에이전트 도구 사용 시 제어 불능 및 반복적인 오류 발생
- 모델 비교 — Qwen이나 Deepseek와 비교했을 때 안정성이 다소 부족함
I was excited to try Mimo given all the buzz on here, so before downloading the quantized local version, I got the token subscription to try it for a month (costs less than a latte). It's really rough, like shockingly bad as some things. A simple prompt that is a layup for every other frontier and local recent model I've tried "write an html page showing a 3d globe", it thought for 10 minutes and came up with this:
Asked it to assume the identity of an apple web designer and critique its prior work and it came up with something much better:
I asked it to make the stars more visible and then it spun out with looping that it could not escape, broke the mouse controls, got fixated on downloading javascript, could not stop using tools when asked not to use tools. I had to break its train of thought by asking it to count back from 20 like a child having a tantrum. And it finally came up with this:
If this were a local quantized model I'd give it a pass, because it handled two other website prompts (a spatial canvas website demo and a pokemon pokedex) reasonably well if somewhat uninspired. It surprisingly passed my Soul Man challenge (asking an LLM who starred in the 1980s comedy Soul Man makes them tend to loop and hallucinate). I'm going to keep trying it out, but Qwen doesn't stumble like this, local Deepseek doesn't stumble like this. Quite strange.
