简而言之 Perplexity 发布了经过训练的 GLM 5.2 版本的研究预览，该版本旨在充当其计算机工具内部的协调器，并仅在需要时升级到 Claude Opus 4.8。 在所有基准测试中，该系统的成本仅为 Opus 4.8 价格的三分之一。 这是 Perplexity 在 18 个月内对中国开源进行的第二次微调，第一个是 R1-1776，这是 DeepSeek R1 的一个版本，删除了大约 300 个北京规定的审查主题。 Perplexity 已将中国的开源模型变成了近乎前沿的主力，其成本约为 Claude Opus 4.8 的三分之一。 该公司今天发布了 Z.AI GLM 5.2 训练后版本的研究预览，该版本是专门为在其计算机代理工具内部运行而构建的，现已投入生产。 我们正在发布 Perplexity Computer 中新编排器模型的研究预览。 该模型是 GLM 5.2 的改编版本，针对计算机线束进行了后训练。它以 Opus 成本的 0.344 倍提供接近前沿的性能。 pic.twitter.com/jcxikoFRfn — 困惑 (@perplexity_ai) 2026 年 7 月 9 日
GLM 5.2 是 Z.ai 的一个大约 7440 亿个参数的模型，前身为智浦人工智能，这是一家北京实验室，自 2025 年 1 月以来一直被列入美国实体清单。（参数是模型在训练过程中可以处理的所有不同的刻度和配置。参数越多，模型就越复杂和强大。）于 6 月在 MIT 许可下发布，它是目前在长期编码基准上可用的顶级人工智能模型之一只需 API 成本的一小部分。 开放权重意味着任何人都可以不受限制地下载、修改和微调它。困惑正是这样做的。 微调实际上是什么 微调是采用已经训练好的人工智能模型并在较小的、集中的数据集上对其进行重新训练的过程，以使其更好地完成特定工作。 把它想象成调整汽车。例如，不同的机械师可以拥有相同的本田思域，使其在飙车中速度更快，视觉上更令人愉悦，使其适应拉力赛等。在人工智能中，开发人员获得一个基本模型并添加不同的设置，以便微调最终获得对特定领域的更多知识、不同的政治偏见、或多或少的限制等。
Perplexity 使用训练后（模型主要训练运行后应用的类似过程）来教授 GLM 5.2 一项关键技能：知道何时处理任务本身以及何时升级到更强大的任务。 这种升级是他们所构建的核心。经过微调的 GLM 5.2 包括 Perplexity 所说的“顾问工具”——一种识别查询何时超出其自身能力并将其移交给第三方前沿模型的本机功能。大多数任务永远不会达到昂贵的模型。只有真正需要的人才会这样做。 这最终节省了大量的推理成本。 首席执行官 Aravind Srinivas 在 X 上写道：“与顾问配合使用时，该模型可实现 Opus 4.8 级性能，而成本仅为一小部分。” 我们一直在对 GLM 的一个版本进行后期训练，该版本经过训练可升级为计算机线束内的前沿模型。当与顾问配合使用时，该模型可以以 Opus 4.8 级的性能运行，而成本仅为其一小部分。现已作为研究预览提供！ — Aravind Srinivas (@AravSrinivas) 2026 年 7 月 9 日
Perplexity 根据正常的 GLM 5.2 对系统进行了基准测试，以建立成本基线。使用公司的内部效率指标来衡量完成复杂任务的成本，结果表明，带有顾问的微调模型的运行成本约为基本版本的两倍。然而，使用顶级 Opus 4.8 型号的所有产品都要昂贵得多（大约贵 600%）。 通过结合这些工具，Perplexity 的系统实现了与 Opus 相同的质量性能，但价格仅为 Opus 的大约三分之一 为何采用中国模式——以及为何开源使其成为可能 中美人工智能竞赛往往被视为零和博弈。在实践中，开源模型并不局限于国界。 GLM 5.2 的 MIT 许可证使计算变得简单：没有需要违反的 API 合同，政府也没有可以翻转的访问开关。您下载权重，然后可以将它们微调到您需要的任何值。
这条路上以前也曾遇到过困惑。当 DeepSeek R1 于 2025 年初席卷人工智能世界时，该公司将其微调为 R1-1776，绘制了大约 300 个由于中国政府审查制度而拒绝讨论的原始主题，并重新训练了模型，使其更加偏向于美国。它成为同一推理引擎的西方托管版本。 Perplexity 的团队当时在一篇博客文章中写道：“如果不先减轻 R1 的偏见和审查制度，我们就无法利用 R1 强大的推理能力。” 因此，GLM 5.2 的举措遵循相同的模板，只不过这次的目标不是政治而是经济。 Perplexity 的计算机产品已经编排了超过 19 个 AI 模型；经过微调的 GLM 被设计为廉价的默认设置，在触及前沿模型之前吸收大量任务。 Srinivas 表示，长期的主题很简单：在已经为数百万用户提供服务的代理工具内，对开源模型进行训练后，以擅长升级。他写道，Perplexity 具有“独特的地位”来解决这个问题，因为基础设施已经大规模部署。
该模型在美国的 Nvidia B200 GPU 上运行。接下来是 Nemotron 3 Ultra 的后期训练，它将使用美国开源模型复制相同的架构。 完整的基准测试和研究论文预计将在未来几周内发布。该模型可作为研究预览。 每日简报时事通讯 每天从当前的热门新闻报道以及原创专题、播客、视频等开始。
In brief
Perplexity released a research preview of a post-trained GLM 5.2 version, built to act as an orchestrator inside its Computer harness and escalate to Claude Opus 4.8 only when needed.
The system costs one-third the price of Opus 4.8 across benchmarks.
It's Perplexity's second Chinese open-source fine-tune in 18 months—the first being R1-1776, a version of DeepSeek R1 stripped of roughly 300 Beijing-mandated censorship topics.
Perplexity has turned a Chinese open-source model into a near-frontier workhorse at roughly a third of what Claude Opus 4.8 costs.
The company released a research preview today of a post-trained version of Z.AI’s GLM 5.2, built specifically to operate inside its Computer agent harness and available now in production.
We're releasing a research preview of a new orchestrator model in Perplexity Computer.
The model is an adapted version of GLM 5.2, post-trained for the Computer harness. It delivers near-frontier performance at 0.344x of the cost of Opus. pic.twitter.com/jcxikoFRfn
— Perplexity (@perplexity_ai) July 9, 2026
GLM 5.2 is a roughly 744-billion-parameter model from Z.ai—formerly Zhipu AI, a Beijing lab that's been on the U.S. Entity List since January 2025. (Parameters are all the different dials and configurations a model can handle during training. The more parameters, the more complex and powerful a model s.) Released under an MIT license in June, it sits among the top AI models currently available on long-horizon coding benchmarks at a fraction of the API cost.
The open weights mean anyone can download, modify, and fine-tune it commercially without restrictions. Perplexity did exactly that.
What fine-tuning actually is
Fine-tuning is the process of taking an already-trained AI model and retraining it on a smaller, focused dataset to make it better at a specific job.
Think of it like tuning a car. Different mechanics can have the same Honda Civic, for example, and make it faster for drag racing, more visually pleasing, adapt it for rally, etc. In AI, developers get a base model and add different settings so the finetune ends up with more knowledge on a specific field, a different political bias, more or less restrictions, etc.
Perplexity used post-training—a similar process applied after the model's main training run—to teach GLM 5.2 one critical skill: knowing when to handle a task itself and when to escalate to something more powerful.
That escalation is the core of what they built. The fine-tuned GLM 5.2 includes what Perplexity calls an "advisor tool"—a native capability to recognize when a query exceeds its own competence and hand off to a third-party frontier model. Most tasks never reach the expensive model. Only the ones that actually need it do.
This ends up saving a lot of money in inference.
"When paired with an advisor, this model functions at Opus 4.8 grade performance at a fraction of the cost," CEO Aravind Srinivas wrote on X.
We’ve been post-training a version of GLM that is trained to escalate to a frontier model inside the Computer harness. When paired with an advisor, this model functions at Opus 4.8 grade performance at a fraction of the cost. Available now as a research preview!
— Aravind Srinivas (@AravSrinivas) July 9, 2026
Perplexity benchmarked the system against the normal GLM 5.2 to establish a cost baseline. Using the company's internal efficiency metric which measures how much it costs to complete complex tasks, the results showed that the fine-tuned model with an advisor is about twice as expensive to run as the basic version. However, using the top-tier Opus 4.8 model for everything is much more expensive (around 600% pricier).
Why a Chinese model—and why open-source makes it possible
The U.S.-China AI race tends to be framed as zero-sum. In practice, open-source models don't stop at borders. GLM 5.2's MIT license makes the calculus simple: There's no API contract to violate, no access switch a government can flip. You download the weights and you can fine-tune them into whatever you need.
Perplexity has been down this road before. When DeepSeek R1 swept through the AI world in early 2025, the company fine-tuned it into R1-1776—mapping roughly 300 topics the original refused to discuss due to Chinese government censorship, and retraining the model to make it more biased in favor of the United States. It became a Western-hosted version of the same reasoning engine.
"We are not able to make use of R1's powerful reasoning capabilities without first mitigating its bias and censorship," Perplexity's team wrote at the time in a blog post.
So, this GLM 5.2 move follows the same template, except the goal this time isn't political but economic. Perplexity's Computer product already orchestrates 19+ AI models; the fine-tuned GLM is designed to be the cheap default that absorbs the bulk of tasks before ever touching a frontier model.
Srinivas said the long-term thesis is straightforward: post-train open-source models to get good at escalation, inside an agent harness that already serves millions of users. Perplexity is "uniquely positioned" to solve it, he wrote, because the infrastructure is already deployed at scale.
The model runs on Nvidia B200 GPUs in the United States. Next in line: a post-train of Nemotron 3 Ultra, which would replicate the same architecture using an American open-source model.
Full benchmarks and a research paper are expected in the coming weeks. The model is available as research preview.
