Cerebras 上的 GLM-4.7：前沿情报达到前所未有的速度

Z.ai 在 Cerebras Inference Cloud 上最新的 GLM-4.7 模型突破了人工智能的界限，为编码和代理任务提供了无与伦比的速度和智能。

GLM-4.7 Unleashed: A Leap in AI Capability on Cerebras

GLM-4.7 释放：Cerebras 人工智能能力的飞跃

The landscape of artificial intelligence is buzzing with the latest announcement from Z.ai: GLM-4.7, the newest iteration in their GLM family of models. Now accessible via the Cerebras Inference Cloud, GLM-4.7 represents a significant stride, merging cutting-edge intelligence with remarkable speed. This powerful combination is poised to revolutionize tasks ranging from complex code generation and editing to sophisticated tool-driven agent interactions and multi-turn reasoning.

Z.ai 的最新公告让人工智能领域充满活力：GLM​​-4.7，这是 GLM 系列模型的最新版本。 GLM-4.7 现在可以通过 Cerebras 推理云访问，代表着重大的进步，以惊人的速度融合了尖端智能。这种强大的组合有望彻底改变从复杂的代码生成和编辑到复杂的工具驱动代理交互和多轮推理等任务。

Frontier Intelligence Redefined

重新定义前沿情报

GLM-4.7 doesn't just improve upon its predecessor, GLM-4.6; it sets a new benchmark. In head-to-head comparisons against leading proprietary models, GLM-4.7 demonstrates a comparable, high-caliber performance in code generation and editing, coupled with robust tool utilization and consistent multi-turn reasoning. What truly sets it apart, however, is its delivery of this intelligence at speeds up to an order of magnitude faster, offering superior price-performance.

GLM-4.7 不仅改进了其前身 GLM-4.6，而且还改进了其前身 GLM-4.6。它树立了新的基准。在与领先的专有模型的正面比较中，GLM-4.7 在代码生成和编辑方面展示了可比的高品质性能，以及强大的工具利用率和一致的多轮推理。然而，真正让它与众不同的是，它以快一个数量级的速度提供这种智能，并提供卓越的性价比。

On benchmarks specifically designed to mirror real-world developer workloads, such as SWEbench, τ²bench, and LiveCodeBench, GLM-4.7 has emerged as the top-ranked open-weight model, surpassing even DeepSeek-V3.2. Developers can expect more accurate, cleaner, and stronger multilingual code outputs. The model exhibits enhanced understanding of project context, improved error recovery, and more stable performance during long, iterative coding sessions.

在专门为反映现实世界开发人员工作负载而设计的基准测试中，例如 SWEbench、τ²bench 和 LiveCodeBench，GLM-4.7 已成为排名第一的开放权重模型，甚至超过了 DeepSeek-V3.2。开发人员可以期待更准确、更清晰、更强大的多语言代码输出。该模型增强了对项目上下文的理解，改进了错误恢复，并在长时间的迭代编码会话中提供了更稳定的性能。

The advancements in GLM-4.7 are also evident in tool-driven agent workflows. The model showcases increased reliability in planning, executing tool calls, and maintaining context across intricate, multi-step interactions. This is largely attributed to its refined internal reasoning capabilities, incorporating an 'interleaved thinking' approach where reasoning precedes each action, and 'preserved thinking,' which allows reasoning context to carry over turns. These enhancements translate to more dependable agents and natural, stable chat and role-play experiences with fewer abrupt shifts in tone or intent.

GLM-4.7 的进步在工具驱动的代理工作流程中也很明显。该模型展示了在复杂的多步骤交互中规划、执行工具调用和维护上下文方面更高的可靠性。这很大程度上归功于其精致的内部推理能力，结合了“交错思维”方法（推理先于每个行动）和“保留思维”，允许推理上下文轮流进行。这些增强功能可以转化为更可靠的代理以及自然、稳定的聊天和角色扮演体验，同时减少语气或意图的突然变化。

Record-Breaking Speed on Cerebras

Cerebras 破纪录的速度

The true game-changer for GLM-4.7 is its ability to operate at real-time speeds on the Cerebras wafer-scale engine. When deployed on Cerebras hardware, GLM-4.7 achieves astonishing code generation speeds of approximately 1,000 tokens per second, and in some instances, up to 1,700 tokens per second. This remarkable feat is made possible by Cerebras's specialized AI hardware, a performance level difficult to match with traditional GPU architectures.

GLM-4.7 真正的游戏规则改变者是它能够在 Cerebras 晶圆级引擎上以实时速度运行。当部署在 Cerebras 硬件上时，GLM-4.7 实现了每秒大约 1,000 个令牌的惊人代码生成速度，在某些情况下，每秒高达 1,700 个令牌。这一非凡的壮举是通过 Cerebras 的专用 AI 硬件实现的，其性能水平难以与传统 GPU 架构相匹配。

This leap in inference speed removes latency as a critical bottleneck, enabling direct deployment into user-facing products and time-sensitive workflows without compromising capability. The real-time performance of GLM-4.7 on Cerebras makes advanced coding assistants, live agents, and latency-sensitive applications not just feasible, but practical, all while maintaining the flexibility of its open-weight design.

推理速度的飞跃消除了延迟这个关键瓶颈，从而可以直接部署到面向用户的产品和时间敏感的工作流程中，而不会影响功能。 Cerebras 上的 GLM-4.7 的实时性能使高级编码助手、实时代理和延迟敏感应用程序不仅可行，而且实用，同时保持其开放式设计的灵活性。

Exceptional Price-Performance

卓越的性价比

While price per token is often considered, the true economic value lies in the speed of useful output. GLM-4.7, running on Cerebras, offers up to a tenfold improvement in price-performance compared to leading closed models like Claude Sonnet 4.5 on demanding coding and agentic tasks. This speed directly translates to reduced end-to-end costs by shortening development cycles, lowering concurrency demands, and minimizing infrastructure requirements for equivalent user experiences.

虽然通常会考虑每个代币的价格，但真正的经济价值在于有用输出的速度。与 Claude Sonnet 4.5 等领先的封闭模型相比，在 Cerebras 上运行的 GLM-4.7 在要求较高的编码和代理任务上的性价比提高了十倍。这种速度可以通过缩短开发周期、降低并发需求以及最大限度地减少同等用户体验的基础设施要求来直接降低端到端成本。

Getting Started with GLM-4.7

GLM-4.7 入门

GLM-4.7 stands as a significant upgrade, solidifying its position as the most potent open-weight model deployed on Cerebras to date. It not only outperforms its open-weight peers but also matches the intelligence of leading closed models in crucial production workloads, all while delivering unparalleled generation speed on Cerebras hardware. For those already using GLM-4.6, migrating is as simple as updating the model name via the API. Z.ai recommends enabling 'preserved thinking' for coding and agentic use cases to maximize benefits.

GLM-4.7 是一次重大升级，巩固了其作为迄今为止 Cerebras 上部署的最有效的开放权重模型的地位。它不仅优于开放重量同行，而且在关键生产工作负载中与领先的封闭模型的智能相匹配，同时在 Cerebras 硬件上提供无与伦比的生成速度。对于已经使用 GLM-4.6 的用户来说，迁移就像通过 API 更新模型名称一样简单。 Z.ai 建议对编码和代理用例启用“保留思维”，以实现利益最大化。

Ready to experience the future? GLM-4.7 is available on the Cerebras Cloud, offering a pay-as-you-go developer tier starting at just $10. This generous tier makes it incredibly easy to prototype, build, and scale your AI applications without hefty upfront investments. Dive in and see what frontier intelligence at lightning speed can do for you!

准备好体验未来了吗？ GLM-4.7 可在 Cerebras Cloud 上使用，提供即用即付的开发人员等级，起价仅为 10 美元。这一慷慨的层级使得您可以非常轻松地原型设计、构建和扩展您的人工智能应用程序，而无需大量的前期投资。深入了解前沿情报以闪电般的速度为您做些什么！