2 ms·
With llamacpp --reasoning-budget truncates thinking and makes it a usefulagain. Its a per client request setting also. Using qwen 3.8 27b, in open code, its u
by cyanydeez 2mo ago
With llamacpp --reasoning-budget truncates thinking and makes it a usefulagain.
Its a per client request setting also.
Using qwen 3.8 27b, in open code, its unnoticeable.
So more a issue for server/harness design than model.
Also, there might be a real bug in the models parameters, like this for 3.8-27b: https://huggingface.co/grimoni/Qwen3.8-27B-SSMFIX-UD-Q4_K_XL-GGUF https://huggingface.co/grimoni/Qwen3.8-27B-SSMFIX-UD-Q4_K_XL...