Open-weight model releases keep stacking up

Three releases hit in one day. Kimi K3.1 Pro weights are out, though commenters noted you need something like an 8x B300 node to run it and that it does not beat DeepSeek V4.1-Flash by much. Baidu released ERNIE 5.5 Flash Base, 280B parameters with 14B active, which one commenter called a chunk better than a DeepSeek model at half the parameter count, but it is only a base model, not tuned to follow instructions. Ripple-65B Preview claims U.S. open-weight SOTA on agentic coding, and the thread mostly complained that the chart ignored the other open models.

A separate Ask HN on replacing Claude or GPT with a local model shows where real usage sits. People report Qwen 3.6 27B dense feeling about like Haiku 4.5, DeepSeek V4 Flash running on 2x RTX Pro 6000 at 160 tok/s, and Gemma 4 on an M4 being too slow. Several said to just plug OpenRouter into your agent and test.

The pattern: open weights are now good enough that the question has moved from whether they work to whether you can afford the hardware to run them. Chinese labs set the pace, and U.S. open-weight efforts are scrambling to be taken seriously.