ByteWarden

glm-5.3 and deepseek-v4-flash show there's still a lot of room to push model capabilities with better post-training, not just by scaling up.

More curious about how qwen3.8-27B performs. That's the size that I can run locally.