v4.57.0

Migration Steps

1. 1. Review the release notes for newly added models: Qwen3-Next, Vault Gemma, Qwen3 VL, Longcat Flash, and Flex Olmo.
2. 2. If you intend to use any of these new models, update your dependency to include the necessary library versions that support them.
3. 3. If you were using older Qwen models, note the introduction of Qwen3-Next, which features architectural changes like Hybrid Attention and High-Sparsity MoE. While direct migration steps aren't specified, be aware of potential performance differences or configuration changes if switching to Qwen3-Next.
4. 4. If you are using Gemma models, be aware of the new Vault Gemma variant, which notably drops norms after Attention and MLP blocks. If you were relying on those norms for stability or specific behavior, you may need to adjust your configuration when switching to Vault Gemma.
5. 5. For users interested in multimodal capabilities, begin exploring the Qwen3-VL series, noting its enhanced MRope and DeepStack integration.
6. 6. For long-context tasks requiring high throughput, consider integrating Longcat Flash, paying attention to its unique shortcut-connected MoE architecture and dynamic parameter activation.