Elorian 首款 AI 模型 Glimpse：要借助图像进行推理，而不只是谈论图像

Elorian 是一家位于帕洛阿尔托的 AI 初创公司，在推出产品之前便以 3 亿美元估值完成了 5500 万美元种子轮融资。TechCrunch 独家获悉，该公司于周五发布了首款模型。

这款名为 Glimpse 的模型以研究预览形式推出，现可通过 API 访问，而聊天界面仍仅限候补名单。因此，目前外界能够看到的主要是一张结果表和一套理念。其理念是什么？当今 AI 模型尽管擅长语言和代码，但仍然无法真正“看见”。

大多数多模态模型只会看一次图像，将其转换为 token，随后以文本形式进行推理——就像凭记忆描述一张照片。Glimpse 则会不断回到原始视觉内容。在推理过程中，它可以裁剪图像中的某个区域、提取视频中的特定帧，或运行代码来测量、比较它正在观察的内容。Elorian 将这一机制称为“视觉草稿空间”（visual scratch space）。这就像有人在手机上双指缩放照片，以再次确认自己刚才看到的内容。

对于曾任 Google DeepMind 研究员、现为 Elorian CEO 的 Andrew Dai 而言，AI 仍无法真正看见的最明显迹象之一，出现在一次游戏之夜。

去年，他拍下了一局刚和朋友结束的桌游，并询问一款前沿模型谁处于领先。模型无法回答。他向 TechCrunch 表示，桌游“各自都有自己的视觉语言、图标体系和符号”。“对成年人来说，即使只有极少文字，也很容易理解发生了什么；但这些模型无法直观理解符号的含义。”

“人类可以看着没有文字的 Ikea 说明图组装 Ikea 家具，对吧？为什么 AI 模型做不到？”他说。

这种方法并非完全新颖——OpenAI 在 2025 年为其 o 系列模型加入了推理过程中的裁剪和缩放能力——但 Elorian 的主张是，它以较低成本围绕这一能力构建了整个模型。该公司表示，一个小团队仅用数百张 GPU 训练了 Glimpse，而且该模型明显小于其在基准测试中对比的前沿系统。

Glimpse 的价格为每百万输入 token 0.60 美元、每百万输出 token 3 美元——低于 Sonnet 和 Sol 等中档模型，但高于 Haiku 和 Luna 等轻量级模型。

不过，根据 Elorian 的评测结果，Glimpse 在科学图表推理和短视频字幕生成方面的表现甚至超过 GPT-6 Astra，并在其他视觉推理任务上取得了具有竞争力的结果。

Elorian 在六个月前结束隐身状态并对外亮相，当时宣布完成一轮融资。本轮融资由 Striker Venture Partners、Menlo Ventures 和 Altimeter Capital 联合领投，Nvidia 和 Google 首席科学家 Jeff Dean 也参与投资。

其最初瞄准的客户，是拥有大量视频数据的企业，包括机器人公司、科学实验室、媒体及创意公司。从长期来看，Elorian 希望从“看见”走向“行动”：利用对场景构建的模型来选择行动方案，包括直接控制机器人。Dai 表示，公司计划在明年打造“世界上最好的视觉推理模型”，即在每一项视觉推理测试中都处于领先地位。

Elorian, the Palo Alto AI startup that raised a $55 million seed round at a $300 million valuation before it launched a product, released its first model on Friday, TechCrunch has exclusively learned.

The model, Glimpse, is launching as a research preview with API access and a chat interface is waitlist-only, so for now it’s a results table and a thesis. The thesis? Today’s AI models, for all their skill with language and code, still can’t really see.

Most multimodal models look at an image once, convert it to tokens, and then reason about it in text — like describing a photo from memory. Glimpse keeps going back to the source. Mid-reasoning, it can crop into a region, pull specific frames from a video, or run code to measure or compare what it’s looking at, a mechanism Elorian calls a “visual scratch space.” Think of someone pinching to zoom on a phone photo to double-check what they thought they saw.

For former Google DeepMind researcher and Elorian CEO Andrew Dai, one of the clearest signs that AI still couldn’t really see came on game night.

Last year he photographed a board game he’d just finished with friends and asked a frontier model who was winning. It couldn’t say. Board games “each have their own visual language, each have their own iconography, symbols,” he told TechCrunch. “To an adult, it’s quite easy to understand what’s going on with minimal text, but these models aren’t able to intuit what the symbols mean.”

“Humans can look at Ikea diagrams, with no text, and assemble Ikea furniture, right? Why can’t AI models?” he said.

The approach isn’t completely new — OpenAI added crop-and-zoom during reasoning to its o-series models in 2025 — but Elorian’s pitch is that it built the whole model around it, cheaply. The company says a small team trained Glimpse on a few hundred GPUs and that the model is substantially smaller than the frontier systems it’s benchmarked against.

Glimpse costs $0.60 per million tokens in and $3 per million tokens out — cheaper than mid-level models like Sonnet and Sol but more expensive than lightweight models like Haiku and Luna.

But according to Elorian’s evaluations, Glimpse outperforms even GPT-6 Astra on scientific chart reasoning and short-form video captioning, with competitive results on other visual reasoning tasks.

Elorian came out of stealth six months ago with a round co-led by Striker Venture Partners, Menlo Ventures, and Altimeter Capital, with participation from Nvidia and Google chief scientist Jeff Dean.

The initial targets are companies sitting on lots of video: robotics firms, scientific labs, media, and creative companies. Longer term, Elorian wants to move from seeing to doing. That means building a model of a scene using that to choose an action, including directly controlling robots. Next year, Dai said, the company aims to build “the world’s best visual reasoning model,” one that leads on every visual reasoning test.