vLLM 团队放出了 d3-turbo，做智能路由用的决策模型，比 d3 flash 快 15 倍，还登顶了 Jev Decision Index，做路由的可以看看。
vLLM 项目宣布其生态新增 d3-turbo，定位为智能路由场景的多模态决策模型。该模型登顶 Jev Decision Index 榜单第一。官方称借助 StateOnce 机制，d3-turbo 比上一代 d3 flash 快 15 倍。
The vLLM ecosystem keeps growing 📷
d3-turbo: SOTA multimodal decision model for intelligent routing — #1 on the Jev Decision Index, 15x faster than d3 flash with StateOnce.
Great work @XunzhuoLiu and the vllm-sr team 📷
Production-grade serving of a MiniMax model through vLLM signals that the open-source inference stack can now carry commercial workloads with SLA guarantees. Teams evaluating cheaper open-weight serving options can look at Inferact's vLLM-based offering as a reference setup. Upstream contributions also mean the wider vLLM community inherits the reliability hardening that came out of this deployment.
MiniMax 新出的 M3.1 Flash，主打代码库工程和多步自动化，vLLM 发布当天就能跑，还带 NVFP4 量化和投机解码，自部署的可以试试。
MiniMax 的 M3.1 Flash 交给了 Inferact 做生产推理，还把优化回馈给 vLLM，用 vLLM 的可以留意
Google 和 Meta 一起做了 TorchTPU，PyTorch 代码不用改就能跑 TPU，vLLM、SGLang 这些库都能配合，已开源在 GitHub。
Liquid AI 推出 RIPL 推理引擎，官方称 LFM2.5-1.2B 的 p95 延迟比 vLLM 跑 Qwen3.5-0.8B 快十倍以上，适合做低延迟应用。
Day-0 support means the model can be self-hosted on existing vLLM infrastructure immediately, without waiting for community integrations. With only 3B active parameters, it targets teams running agentic workloads on modest local hardware who want an alternative to larger dense models. Developers comparing open-source MoE options for on-premise serving should evaluate its throughput against Qwen and DeepSeek MoE checkpoints.
