多元开放模型发布：Cohere Command A+、NVIDIA Nemotron-3-Ultra、Zyphra ZAYA1、Poolside Laguna-M.1 等
英文摘要
The open model ecosystem continues to diversify, with multiple organizations releasing new models this week. Cohere's Command A+ (218B-A25B MoE) is now under Apache 2.0 license and offers multimodal, multilingual, and agentic capabilities, usable on a single B200 GPU with 4-bit quantization. NVIDIA released Nemotron-3-Ultra-550B-A55B-BF16 under the OpenMDW license, a LatentMoE model that is faster than comparables. Other notable releases include Zyphra's ZAYA1-74B-preview MoE model, Poolside's Laguna-M.1 under Apache 2.0 (with a commitment to future open releases), and updated models from Zhipu (GLM-5.2), Moonshot AI (Kimi-K2.7-Code), and Stepfun (Step-3.7-Flash, strong in math). This breadth shows a shift from a few dominant players to a global, multi-actor open model landscape.
中文摘要
开放模型生态持续多元化，本周多家机构发布新模型。Cohere 的 Command A+（218B-A25B MoE）已采用 Apache 2.0 许可，具备多模态、多语言和智能体能力，可在单张 B200 GPU 上以 4-bit 量化运行。NVIDIA 以 OpenMDW 许可发布了 Nemotron-3-Ultra-550B-A55B-BF16，这款 LatentMoE 模型速度优于同类模型。其他亮点包括 Zyphra 的 ZAYA1-74B-preview MoE 模型、Poolside 的 Apache 2.0 许可的 Laguna-M.1（并承诺未来持续开放发布），以及智谱（GLM-5.2）、月之暗面（Kimi-K2.7-Code）和阶跃星辰（Step-3.7-Flash，数学能力强）的更新模型。这种广度表明开放模型格局已从少数主导者转向全球性的多参与者生态。
关键要点
Cohere released Command A+ (218B-A25B MoE) under Apache 2.0 license, with multimodal, multilingual, and agentic capabilities, runnable on a single B200 GPU at 4-bit.
Cohere 以 Apache 2.0 许可发布了 Command A+（218B-A25B MoE），具备多模态、多语言和智能体能力，可在单张 B200 GPU 上以 4-bit 运行。
NVIDIA launched Nemotron-3-Ultra-550B-A55B-BF16, a LatentMoE model faster than peers, under the model-weight-specific OpenMDW license.
NVIDIA 发布了 Nemotron-3-Ultra-550B-A55B-BF16，采用专为模型权重设计的 OpenMDW 许可，其 LatentMoE 架构速度超越同类模型。
Zyphra's ZAYA1-74B-preview (74B-A4B MoE) and Poolside's Laguna-M.1 were both open-sourced under Apache 2.0, with Poolside committing to open weights as default going forward.
Zyphra 的 ZAYA1-74B-preview（74B-A4B MoE）和 Poolside 的 Laguna-M.1 均以 Apache 2.0 许可开源；Poolside 承诺未来将开放权重作为默认策略。
Zhipu's GLM-5.2 remains competitive for everyday use, while Kimi-K2.7-Code (token-efficient), Step-3.7-Flash (strong math), and Nemotron-Labs-Diffusion-14B (supports autoregressive, diffusion, and self-speculation modes) also saw updates.
智谱的 GLM-5.2 在日常使用中保持竞争力，同时 Kimi-K2.7-Code（注重令牌效率）、Step-3.7-Flash（数学能力强）和 Nemotron-Labs-Diffusion-14B（支持自回归、扩散和自推测三种模式）也获得更新。