⚡ The main thing in 60 seconds (TL;DR)
⏱ ~ 2 мин чтения- The larger context simply allows me to do more work instead of having to re-read ETC ID files, and rather use my extra resources in the context window. Instead of more parameters. I get my 400 thousand ctx q...
- I get my 400k ctx qwen3.8-27b + internet search, get the latest up-to-date documents due to lack of world knowledge, and I don't spend extra Vram storing data I don't need.
- Additionally, more active settings allow the film to use a broader mental model of the task.
The larger context simply allows me to do more work instead of having to re-read ETC ID files, and rather use my extra resources in the context window. Instead of more parameters.
I get my 400k ctx qwen3.8-27b + internet search, get the latest up-to-date documents due to lack of world knowledge, and I don't spend extra Vram storing data I don't need. Additionally, more active settings allow the film to use a broader mental model of the task. The Qwen4-27b will be designed to make 1M more reliable, and using it for context on a lower density model seems more productive to me. Any. Also note that the agent session actually works when ctx is higher? This is a phenomenon I've encountered, I tried 3.8flash-next, yes it's great and much faster, but I prefer the reliability of that big context window load.
In particular, you have a spare vram, if you can get 1M on it and have a large MOE with at least 12b active options + 1M, it would be better for a MOE lot. Much better than qwen4-27b, but at 48GB vram + 64GB system memory I think it's more efficient for a smaller model + 1M makes more sense
💬 Discussion 0
BBCode & Emoji supportedParticipate in discussions, ask authors questions, and share benchmarks by joining the ProArena engineering community.