I'm a half-farmer,
Sometimes I make and run Qwen 3.8 FN while continuing to use it for my main job, but also as a hobby, and my company juniors who are doing the same thing don't understand.
Today, I was asked to make a report on the future direction based on internet and paper standards.
GPT is free so I used 5.6 Luna. It took 4-5 minutes to complete. Just normal.
Qwen 3.8 FN seems to be a feature. It takes 1 hour to make + 1 hour to refine -_-;;; and another 2 hours.
The result is...;;; It seems like it usually makes about 20 pages.
The writing seems to be a bit condensed, but it specifies the source, purpose, and relationship with the repository I instructed, and even plans for directionality. It's as good as GLM 5.3 Flash, which is used in our company.
And when I threw it to GPT and said it was made by a local llm, I came up with all sorts of excuses and finally admitted defeat ㅎㅎㅎ
Even in the company, they miss Qwen 3.8 FN. When I asked, they said it wasn't possible because of the license policy.
(Next model seems to be something different)
The llamacpp tuning for StrixHalo is also quite stable, and I'm currently expanding the work based on reports by continuously filling 256K, compacting, and repeating.
