小米于9月22日凌晨发布并开源了Xiaomi MiMo-V2.6系列模型，包含Pro与Flash两个原生全模态版本1。其中MiMo-V2.6-Pro在AA综合智能指数v4.3.2中获得46分，超越Kimi K3的44分和GLM-5.3的45分，成为当前AA指数排名最高的开源权重模型1。相比前代产品，MiMo-V2.6-Pro的得分从26分上升至46分1。
小米通过规模化扩展强化学习算力来优化模型性能1。MiMo-V2.6-Pro和Flash的训练成本分别约为262万美元和85万美元，训练耗时均不到6天1。在DeepSWE v1.1评测中，Pro模型性能从48.8提升至65.7，Flash模型则从58.4提升至72.61。此外，两个版本在训练任务上的平均通过率分别相对提升了25%和12%1。
定价方面，Pro与Flash版本保持与前代一致的定价体系1。Flash提供每百万词元输入1元、输出2元的价格，Pro则为输入3元、输出6元，两版本均提供99%缓存折扣1。该系列模型支持100万上下文长度训练，单次强化学习训练使用1568个样本进行更新1。小米已全面开源MiMo-V2.6-Pro和Flash的模型权重、技术报告及强化学习研究资源1。
Xiaomi unveiled and open-sourced its new MiMo-V2.6 series of multimodal AI models in the early hours of September 22, comprising both Pro and Flash variants 1. The MiMo-V2.6-Pro achieved a score of 46 points on the AA Composite Intelligence Index v4.3.2, surpassing Kimi K3 (44 points) and GLM-5.3 (45 points) to become the highest-ranking open-source weights model on the benchmark 1. This represents a 20-point improvement over its predecessor, MiMo-V2.5-Pro, which scored 26 points 1.
Leveraging scaled reinforcement learning compute, both models were developed with significant efficiency gains 1. Training costs for the Pro and Flash versions reached approximately $2.62 million and $850,000 respectively, with training completed in under six days 1. On the DeepSWE v1.1 evaluation, the Pro model improved from 48.8 to 65.7, while Flash advanced from 58.4 to 72.6 1. The models support context lengths of up to one million tokens and employ 1,568 samples per single RL training update 1. Training task pass rates improved by 25% for Pro and 12% for Flash relative to previous versions 1.
Pricing for the two variants remains unchanged: Flash is priced at 1 yuan per million input tokens and 2 yuan per million output tokens, while Pro costs 3 yuan and 6 yuan respectively, with a 99% cache discount available 1. Xiaomi has made model weights, technical reports, and RL research resources fully open-source 1.
评论
还没有评论，欢迎留下第一条。