Qwen 3.8 Max (2.4T, 1bit): 53,448 tok, 1h 54m, 7.8 tok/s
Ornith 1.5 (397B): 53,365 tok, 38m, 23.5 tok/s GLM 5.2: 20,864 tok, 12m, 29.6 tok/s
Relay Reasoning improved verified accuracy in my local experiments by carrying forward useful results and execution feedback between attempts.
Qwen 3.8 27B, local: MATH-500: 63.6% → 89.0% AIME 2026: 60.0% → 70.0%
Ornith 1.5 35B and Qwen 3.8 27B both dropped this month promising the best local AI model you can run at home.
Only one of them has actually proven it. This breakdown puts ornith_ 1.5 35B vs Alibaba_Qwen 3.8 27B head to head using real benchmarks, independent testing, and
Who says Beggar’s Edition Mac can’t run large models? 16G Beggars Edition M1 deployment Qwen 3.8-27B code actual test
Not surprisingly, the IQ is online, Opus-4.6 level! ! ! Well-deserved reputation The industry-famous "Pelican Cycling" test can be easily passed with eyes closed 🚴♂️ Although the Token output speed is slightly slower, the overall operation is flawless. Friends with Macs, boldly deploy, Token freedom is not a dream! Big ups to…
You have reached the end of the archive
All of qwen38