美国开源模型为何落后 Mistral 4 差距有限
有人吐槽美国开源模型进展慢:Mistral 4 看着落后,其实和 Llama 那些美国开源模型差不了多少。
一条推文指出所谓美国开源权重模型的复兴并未出现。作者认为 Mistral 4 虽被嘲笑落后于前沿水平,但在 Artificial Analysis 排名上与美国最好的开源模型差距在误差范围内。作者预计后训练环节将得到强化。
Not really seeing "American open weight Renaissance" yet Many have laughed at Mistral 4 for being so far behind the frontier, but Mistral rn is like, within the margin of error from the best American open model (at least on AA). I really expect post-training to be reinforced https://t.co/IFOwn9hclk