内容来源：https://aibusiness.com/generative-ai/openai-gpt-5-3-codex-spark-shows-what-s-possible-with-cerebras

内容总结：

OpenAI发布新型编码模型，首次采用非英伟达AI芯片

2月12日，OpenAI发布了新一代编码模型GPT-5.3-Codex-Spark的研究预览版。该模型专注于实时编程辅助，是其本月早前发布的GPT-5.3-Codex的轻量化版本，旨在为开发者提供即时、低成本的代码生成与修改支持。

此次发布备受业界关注的核心原因在于其底层硬件架构的转变：Codex-Spark是OpenAI首个完全未使用英伟达（Nvidia）硬件、转而搭载初创公司Cerebras的Wafer-Scale Engine 3芯片运行的模型。这一合作被视作AI芯片市场格局可能出现变化的一个信号。

分析指出，此举对合作双方均具有战略意义。对Cerebras而言，成功支撑OpenAI模型有助于向企业客户证明，其大尺寸AI芯片及晶圆级设计能够与市场主导的英伟达GPU竞争。对OpenAI来说，在主要竞争对手Anthropic近期获得巨额融资并加大市场投入的背景下，持续优化其编码模型有助于巩固其在企业级市场的竞争力。

行业分析师苏连杰评价称，Codex-Spark作为一款轻量化模型，在支持实时编码、降低开发成本方面具有优势，尤其适合初学者或需要即时辅助的程序员。但其功能也存在一定局限，例如仅支持文本提示，上下文窗口为128k。

尽管模型目前表现聚焦于特定应用场景，但OpenAI与Cerebras的合作若取得长期成功，可能为其他AI硬件供应商（如专注于专用集成电路的厂商）打开市场空间，促使更多AI模型制造商考虑多元化的硬件选择。不过，从英伟达生态转向其他硬件平台，仍需在后台工程层面进行大量重构与适配工作。

对企业用户而言，技术路线的竞争最终将回归产品本身的实际效能。模型的准确性、响应速度以及是否真正实现低延迟，才是他们评估采纳与否的关键标准。

中文翻译：

由谷歌云赞助
选择您的首个生成式AI应用场景
要开始使用生成式AI，首先应关注能够提升人类信息交互体验的领域。

该供应商专为实时编码设计了此模型。虽然功能有限，但它展现了当AI供应商选择英伟达之外的硬件合作伙伴时可能带来的创新空间。

在争夺企业级市场的竞争中，OpenAI推出了基于初创公司Cerebras先进AI芯片驱动的新型编码模型。

2月12日，OpenAI以研究预览版形式发布了新型编码模型GPT-5.3-Codex-Spark。这家生成式AI供应商将该模型设计为GPT-5.3-Codex的精简版，专注于实时编码应用。而完整版GPT-5.3-Codex已于本月初发布，适用于计算机操作与编码任务。

Codex-Spark成为首个未采用英伟达硬件的OpenAI模型，完全依托Cerebras晶圆级引擎3芯片运行。

此次发布正值Cerebras与OpenAI共同向企业客户证明其超越竞争对手的价值之际。

对专注于半导体与云计算的初创企业Cerebras而言，Codex-Spark的效能可向潜在客户证明：其大型AI芯片与晶圆级设计能够匹敌英伟达占据市场主导地位的GPU。与此同时，在Anthropic主导的市场中，OpenAI通过提升编码模型性能可增强自身竞争力——尤其考虑到热门模型Claude的创造者刚获得300亿美元新融资，并投入2000万美元成立超级政治行动委员会以抗衡OpenAI的同类组织。

GPT-5.3-Codex-Spark高度聚焦实时编码场景。OpenAI表示该模型可与Codex协同完成定向编辑、逻辑重构及即时任务处理。

Omdia（Informa TechTarget旗下机构）分析师苏炼杰指出，作为轻量化模型，Codex-Spark更易于部署且能为寻求即时支持的开发者提供更高性价比，但部分功能存在局限。例如其上下文窗口仅128K且仅支持纯文本提示。

"对于某些应用场景，这或许已足够用。该模型精准定位特定编码人群的设计确实巧妙，"苏炼杰补充道，初学者或需要实时编码辅助的用户可能会发现Codex-Spark最具吸引力。

此外，OpenAI采用Cerebras晶圆级引擎基础设施的举措，可能为其他专注于专用集成电路（ASIC）的AI硬件供应商（如Grok或Tenstorrent）创造市场机遇。

"通过提供基于AI ASIC（特别是专为推理设计的芯片）的服务，为市场同类参与者开辟了新的商业模式，"苏炼杰强调。他补充说明，借助Cerebras的高吞吐量推理芯片，Codex-Sparks能够实现低延迟实时推理。

但值得注意的是，OpenAI使用Cerebras替代英伟达GPU仍需在后端进行大量工程配置。苏炼杰指出，从GPU转向晶圆级引擎需要突破英伟达架构体系，进行大规模重新配置、移植及代码库转换。

不过他也表示，若OpenAI与Cerebras的合作取得成功，未来其他AI模型制造商或将尝试更多硬件选择方案。

对企业用户而言，系统底层硬件的意义远不及产品实际效能。

"他们只关心产品是否适用——准确性、响应速度是否真如宣传所言达到超低延迟，"苏炼杰总结道。

英文来源：

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
The vendor designed the model for real-time coding. While limited, it shows the possibilities that can arise when an AI vendor chooses a hardware vendor other than Nvidia.
As it battles rival Anthropic for the enterprise market, OpenAI introduced a new coding model powered by an advanced AI chip from startup Cerebras.
OpenAI released the new coding model, GPT-5.3-Codex-Spark, in research preview on Feb. 12. The generative AI vendor designed the model for real-time coding as a smaller version of GPT-5.3-Codex, which OpenAI released earlier this month for computer use and coding tasks.
Codex-Spark is the first OpenAI model not to use Nvidia’s hardware, running solely on Cerebras Wafer-Scale Engine 3 chips.
The release of Codex-Sparks comes as both Cerebras and OpenAI are trying to prove to enterprises their worth over their competitors.
For Cerebras, a specialized semiconductor and cloud computing startup, Codex-Spark's effectiveness could show potential customers that its large AI chips and wafer-scale design can be just as valuable as Nvidia’s market-dominating GPUs. Meanwhile, for OpenAI, improving its coding models could boost its credibility in a market dominated by Anthropic, especially given that the creator of the popular Claude model just raised another $30 billion and is putting $20 million into a new super PAC to counter OpenAI's super PAC.
GPT-5.3-Codex-Spark focuses heavily on real-time coding. OpenAI said the model works with Codex for tasks such as targeted edits, reshaping logic, and getting work done in the moment.
As a small model, Codex-Spark is easier to support and more cost-efficient for developers looking for immediate support; however, it is limited in some features, said Lian Jye Su, an analyst at Omdia, a division of Informa TechTarget. For example, the context window is only 128k and supports text-only prompts.
“For some use cases, it's probably sufficient, and it's really cleverly designed to target a specific segment of the coding population,” Su said. He added that beginner coders or those seeking real-time coding assistance might find Codex-Spark most appealing.
Moreover, OpenAI's use of Cerebras’ wafer-scale engine infrastructure could represent an opportunity for other AI hardware vendors, such as Grok or Tenstorrent, that specialize in application-specific integrated circuits or ASICs.
“By making these services available and running on AI ASICs, particularly the ones that are specifically designed for inference, creates this business model for similar players in the market,” Su said. He added that, with Cerebras’ high-throughput inference chips, Codex-Sparks supports low-latency, real-time inference.
It is clear, however, that much of the engineering still needs to be configured on the backend for OpenAI's use of Cerebras GPUs rather than Nvidia's. Su added that the shift from GPU to the wafer-scale engine will require significant reconfiguration, porting, and codebase conversion outside Nvidia’s architecture.
However, if OpenAI succeeds with Cerebras, other AI model makers could decide to try other hardware options in the future, Su added.
For enterprises, the hardware behind the system matters less than whether the product functions effectively.
“They only care about whether this works for them or not, the accuracy, the responsiveness, whether it is really as they said, like really low latency,” he said.