kisjovan

3 ms·

1. 1.▲
   by kisjovan 9d ago

   There is! We've got 20+ community quants
2. 2.▲
   by kisjovan 9d ago

   It does, but very rarely (as opposed to the Base Qwen 3.8 27B), making it actually usable for day-to-day work - which is why we even trained it
3. 3.▲
   by kisjovan 9d ago

   Did you get to try it?
4. 4.▲
   by kisjovan 9d ago

   Thank man appreciate it! Did you get the chance to try it?
5. 5.▲
   by kisjovan 9d ago

   Thank you so much for trying it! I personally haven't tried it - the community has noted that it does indeed work though. On the Swift Qwen3.8-Flash-Next, we're running the benchmarks right now and will get it out end of this or s
6. 6.▲
   by kisjovan 10d ago

   You should check out our TerminalBench2.1 score for that, the tasks there can run up to 4h! We got almost no loss (we got Base 66.74% vs Swift 65.84%) with -38.7% thinking token reduction. There's also quite a few independent evals on
7. 7.▲
   by kisjovan 10d ago

   Thank you so much!! It would be great if you could share some numbers with the community :)
8. 8.▲
   by kisjovan 10d ago

   I will TLDR you on our thought process, research, training and benchmarks. 1. When running our quantized Qwen 3.8 27B instances we were very annoyed by random reasoning loops (in the paper bellow refered to as "overthinking errors&quot
9. 9.▲
   Show HN: Swift-Qwen3.8-27B, -58.3% thinking, x1.95 speed, accuracy of xhigh(huggingface.co)

   13 points by kisjovan 10d ago|5 comments