Liquid AI 发布 d1 决策模型并新增图像输入能力

来源:Liquid AI 模型与工程博客(网页) 2026年10月5日 08:00 AIHOT 评分:69 精选
摘要:Liquid AI 发布 d1 决策模型,新增文本与图像输入,可通过 console.liquid.ai 和 d1 Playground 使用。
推荐理由:原文给出 d1 在六类真实应用中对 GPT-6.1 Sol 和 Claude Opus 5.5 的成本与速度对比,以及按输入 token 计费规则,便于评估是否替换现有 LLM 调用。
# Liquid AI 发布 d1 决策模型并新增图像输入能力 > 原标题:Introducing d1: The most capable decision model, now with vision - 来源:Liquid AI 模型与工程博客 - 发布时间:2026-10-05T00:00:00.000Z - AIHOT:https://aihot.news/items/uu1qa3hh83kc9i4u832wpywyp - 原文:https://www.liquid.ai/blog/d1-decision-model ## 摘要 Liquid AI 发布 d1 决策模型,新增文本与图像输入,可通过 console.liquid.ai 和 d1 Playground 使用。 ## 推荐理由 原文给出 d1 在六类真实应用中对 GPT-6.1 Sol 和 Claude Opus 5.5 的成本与速度对比,以及按输入 token 计费规则,便于评估是否替换现有 LLM 调用。 ## 正文 · 中文译文 今天,我们推出 **d1,我们的首个决策模型**,现已支持文本和图像。通过[上周的实验性发布](https://x.com/liquidai/status/2105003472332693869),d1 成为首个在文本决策上可与 Jev 匹敌的模型,如今它又率先将这些能力扩展到图像。你今天就可以在[console.liquid.ai](http://console.liquid.ai/) 和 [d1 Playground](http://d1.liquid.ai/) 中试用。 我们在六个真实应用中,将 d1 与 GPT-6.1 Sol 和 Claude Opus 5.5 进行了对比测试,从筛选支持工单到检查电路板。在其中四个应用上,d1 与 GPT-6.1 Sol 持平或更优。它的成本比这两个模型低 19 倍到 200 倍,并且在每项任务上都显著更快。 ## d1 决策模型如何工作 决策模型针对某个情境,为每个可能答案给出一个概率来回答问题。d1 决策模型接收非结构化数据(例如文本、图像或两者)以及一个或多个问题作为输入,在一次前向传播中读取它们,并返回概率,而不生成任何 token。一次文本决策耗时 200 到 300 毫秒,快到足以用于实时应用。 d1 回答三类问题: - **Noul**:一个是/否问题,用 0 到 1 之间的概率作答。 - **Choice**:从多个标签中选一个,为每个标签给出一个概率。 - **Score**:在一个量表上的位置,由每个等级的概率加权。 一个请求可以针对同一状态提出多个问题,从而节省输入 token。 ## d1 实战 当答案是结构化决策时,决策模型可以替代对语言模型的高成本调用。它们还催生了各种新用例,本节展示了其中一些。下面的每个演示都在 [d1 Playground](http://d1.liquid.ai/) 中实时运行。 **视觉检测。**这个工业应用检查来自四条产线、经过摄像头下方的零件:电路板、蜡烛、腰果和口香糖(公开的 VisA 数据集)。d1 以 85-97% 的准确率分拣良品和缺陷品。最有趣的是,该模型从未为此训练过。得益于其出色的泛化能力,它能从一段简短描述中理解任务。 **文本应用。**五个应用将 d1 用于每一个决策: - **Smart Filter:**d1 成为 SQL 查询中的一个函数,`WHERE d1(ticket, 'the customer wants to cancel)`。它为 150 个支持工单逐一回答是或否。 - **Code Search:**d1 逐层浏览 Hugging Face transformers 仓库(6,511 个文件),一次一个文件夹,直到找到能回答问题的那个函数。 - **Smart Folders:**d1 将每个新文档归入一个文件夹,再归入一个子文件夹。它以同样方式归档搜索问题,因此关键词搜索只需查看一个子文件夹。 - **Web Agent:**d1 根据一句话目标操作一个航班搜索网站。每一步,它都会在页面允许的所有操作中选出下一个动作。 - **Context Compaction:**d1 读取编码智能体会话中的每个工具输出,并为下一个任务保留、裁剪或丢弃它。它移除了 52% 的 token,同时保留了任务所需的每一个输出。 我们改编了其中四个应用,来自以下开源项目: [pg-jev](https://github.com/realZachi/pg-jev)、 [jevgrep](https://github.com/dzhng/jevgrep)、 [jev-ultrafast](https://github.com/browser-use/jev-ultrafast) 和 [fast-jev-compaction](https://github.com/tamaratran/fast-jev-compaction)。 **游戏。**d1 实时玩八款经典游戏,并选择每一步。视觉在两方面帮助它: - **更优的决策**:俄罗斯方块完全可以用文字描述,但加入屏幕画面后,d1 的得分从消除 70 行提升到 81 行。 - **更简单的集成**:在 Wordle 中,d1 直接从截图读取棋盘。开发者无需编写游戏的文字版本,而 d1 仍以平均 3.8 次猜测解决了 12 局中的 12 局。 d1 也能处理纯视觉任务。在 Quick, Draw! 中,它能在 62 个词中猜出玩家的涂鸦内容,并能识别 6 幅画中的 5.2 幅(随机猜测为 0.6)。 ## 可用性与定价 立即使用 d1 决策模型开始构建,该模型已在 [Liquid AI API](http://console.liquid.ai/) 上以 `d1` 形式提供。在 [console.liquid.ai](https://console.liquid.ai/) 创建 API 密钥(Dashboard > API Keys)。 要使用视觉功能,您只需在 `images` 中以 base64 数据 URL 的形式发送图像: ``` import base64, os, requests image = base64.b64encode(open("board.jpg", "rb").read()).decode() response = requests.post( "https://api.liquid.ai/decisions/v1/systemone", headers={"Authorization": f"Bearer {os.environ['LIQUID_API_KEY']}"}, json={ "model": "d1", "images": [f"data:image/jpeg;base64,{image}"], "state": "Camera image of a circuit board on the production line.", "questions": {"defect": { "type": "noul", "instructions": "Does this circuit board have a defect?" } }, }, ) print(response.json()["answers"]["defect"]["noul"]) ``` 完整的 API 参考文档见 [决策模型文档](https://docs.liquid.ai/lfm/models/decision-models)。 d1 仅按输入 token 计费,无输出 token。图像按与文本相同的费率计为输入 token:每 32×32 像素块 1.5 个 token,因此一张 1024×1024 的图像消耗 1,536 个 token。每个问题作为独立的提示计费,包括其文本和所有图像。 d1 也可通过 [Vercel](https://vercel.com/ai-gateway/models/d1) 和 [OpenRouter](https://openrouter.ai/liquid/d1) 使用,目前仅支持文本。视觉功能即将登陆这两个平台。 这是 AI 创造与探索的激动人心的时刻,而 d1 仅仅是开始。我们正在继续开发各种规模的决策模型,带来新功能、更高的决策质量和更低的延迟,并计划很快在 Hugging Face 上发布即将推出的模型的开源权重。 [![Liquid AI logo](https://www.liquid.ai/_next/image?url=https%3A%2F%2Faypchzzf9pftwuto.public.blob.vercel-storage.com%2Fliquid-logo-xIjdtFz2499igEZp35jOGE8qF8wpBA&w=48&q=75)在 Liquid Console 上试用](https://console.liquid.ai/)[![Liquid AI logo](https://www.liquid.ai/_next/image?url=https%3A%2F%2Faypchzzf9pftwuto.public.blob.vercel-storage.com%2Fliquid-logo-xIjdtFz2499igEZp35jOGE8qF8wpBA&w=48&q=75)阅读我们的文档](https://docs.liquid.ai/lfm/models/decision-models) ## 引用 如需引用,请使用以下参考文献或 BibTeX: Liquid AI,“Introducing d1: The most capable decision model, now with vision”,Liquid AI Blog,2026 年 10 月。 方法说明。我们于 2026 年 10 月 5 日使用 d1 Playground 的对比脚本对每个模型运行了每个应用一次。GPT-6.1 Sol 和 Claude Opus 5.5 将每个请求作为一条聊天消息接收并以 JSON 格式回答,采用其默认推理设置。当多个问题共享一个输入时,例如过滤器的工单或文件夹的段落,聊天模型会分批回答。成本使用标价,不含提示缓存折扣。d1 的成本使用每百万输入 token 0.04 美元。时间为每次运行的时间,最多 8 个请求并发。一次 Smart Filter 运行是对 150 个工单的一次查询。一次 Smart Folders 运行归档 105 个段落,其成本按每 1,000 个段落计算。过滤器的质量是其相对于人工标注的 F1 分数,该分数同时计入遗漏和错误的匹配。其他文本应用统计正确处理的目标、问题、段落或所需输出。我们在 d1 的流程设定后编写了 15 个代码问题中的 6 个以及 4 个压缩会话中的 2 个。在视觉检查中,每个模型都会看到来自同一产线的良品部件与待检查部件并排展示。 ## 正文 · 原文 Today, we introduce **d1, our first decision model**, now supporting both text and images. With [last week's experimental release](https://x.com/liquidai/status/2105003472332693869), d1 became the first model to rival Jev on text decisions, and it is now the first to extend these capabilities to images. You can try it today at[console.liquid.ai](http://console.liquid.ai/) and in the [d1 Playground](http://d1.liquid.ai/). We tested d1 against GPT-6.1 Sol and Claude Opus 5.5 on six real applications, from filtering support tickets to inspecting circuit boards. d1 matches or beats GPT-6.1 Sol on four of them. It costs 19x to 200x less than both models and answers significantly faster on every task. ## How the d1 decision model works Decision models answer questions about a situation with a probability for each possible answer. The d1 decision model takes unstructured data (e.g., text, images, or both) and one or more questions as input, reads them in one forward pass, and returns the probabilities, without generating any tokens. A text decision takes 200 to 300 ms, fast enough for real-time applications. d1 answers three types of questions: - **Noul**: a yes/no question, answered with a probability between 0 and 1. - **Choice**: pick one label among many, answered with a probability per label. - **Score**: a position on a scale, weighted by the probability of each level. One request can ask several questions about the same state, saving input tokens. ## d1 in action Decision models can replace expensive calls to language models when the answer is a structured decision. They also enable a variety of new use cases, some shown in this section. Every demo below runs live in the [d1 Playground](http://d1.liquid.ai/). **Visual inspection.** This industrial application reviews parts from four production lines that pass under a camera: circuit boards, candles, cashews, and chewing gum (public VisA dataset). d1 sorts good and defective parts with 85-97% accuracy. The most interesting part is that the model was never trained for this. Thanks to its excellent generalizability, it understands the task from a short description. **Text applications.** Five applications use d1 for every decision: - **Smart Filter:** d1 becomes a function in a SQL query, `WHERE d1(ticket, 'the customer wants to cancel)`. It answers yes or no for each of 150 support tickets. - **Code Search:** d1 browses the Hugging Face transformers repository (6,511 files) one folder at a time, down to the function that answers a question. - **Smart Folders:** d1 files each new document into a folder, then a subfolder. It files search questions the same way, so keyword search only looks in one subfolder. - **Web Agent:** d1 operates a flight-search website from a one-sentence goal. At each step, it picks the next action among everything the page allows. - **Context Compaction:** d1 reads each tool output in a coding agent's session and keeps, trims, or drops it for the next task. It removes 52% of the tokens and keeps every output the task needs. We adapted four of these applications from the following open-source projects: [pg-jev](https://github.com/realZachi/pg-jev), [jevgrep](https://github.com/dzhng/jevgrep), [jev-ultrafast](https://github.com/browser-use/jev-ultrafast), and [fast-jev-compaction](https://github.com/tamaratran/fast-jev-compaction). **Games.** d1 plays eight classic games live and picks every move. Vision helps it in two ways: - **Better decisions**: Tetris can be fully described in text, but adding the screen raises d1's score from 70 to 81 cleared lines. - **Simpler integration**: In Wordle, d1 reads the board directly from a screenshot. Developers don't need to write a text version of the game, and d1 still solved 12 of 12 games in 3.8 guesses on average. d1 also handles purely visual tasks. In Quick, Draw!, it guesses what a player's doodle shows among 62 words, and recognizes 5.2 of 6 drawings (random guessing gets 0.6). ## Availability and pricing Start building today with the d1 decision model, available on the [Liquid AI API](http://console.liquid.ai/) as `d1`. Create an API key at [console.liquid.ai](https://console.liquid.ai/) (Dashboard > API Keys). To use vision, you can simply send images as base64 data URLs in `images`: ``` import base64, os, requests image = base64.b64encode(open("board.jpg", "rb").read()).decode() response = requests.post( "https://api.liquid.ai/decisions/v1/systemone", headers={"Authorization": f"Bearer {os.environ['LIQUID_API_KEY']}"}, json={ "model": "d1", "images": [f"data:image/jpeg;base64,{image}"], "state": "Camera image of a circuit board on the production line.", "questions": {"defect": { "type": "noul", "instructions": "Does this circuit board have a defect?" } }, }, ) print(response.json()["answers"]["defect"]["noul"]) ``` The full API reference is in the [decision models documentation](https://docs.liquid.ai/lfm/models/decision-models). d1 is billed on input tokens only, with no output tokens. Images are counted as input tokens at the same rate as text: 1.5 tokens per 32×32-pixel patch, so a 1024×1024 image costs 1,536 tokens. Each question is billed as its own prompt, including its text and all images. d1 is also available through [Vercel](https://vercel.com/ai-gateway/models/d1) and [OpenRouter](https://openrouter.ai/liquid/d1), with text only for now. Vision is coming to both soon. This is an exciting time for creativity and exploration in AI, and d1 is only the beginning. We're continuing our work on decision models across a range of sizes, with new features, higher decision quality, and lower latency, and we plan to release open weights for upcoming models on Hugging Face soon. [![Liquid AI logo](https://www.liquid.ai/_next/image?url=https%3A%2F%2Faypchzzf9pftwuto.public.blob.vercel-storage.com%2Fliquid-logo-xIjdtFz2499igEZp35jOGE8qF8wpBA&w=48&q=75)Try on Liquid Console](https://console.liquid.ai/)[![Liquid AI logo](https://www.liquid.ai/_next/image?url=https%3A%2F%2Faypchzzf9pftwuto.public.blob.vercel-storage.com%2Fliquid-logo-xIjdtFz2499igEZp35jOGE8qF8wpBA&w=48&q=75)Read our docs](https://docs.liquid.ai/lfm/models/decision-models) ## Citation For citations, please use the following reference or BibTeX: Liquid AI, "Introducing d1: The most capable decision model, now with vision", Liquid AI Blog, Oct 2026. Methodology. We ran each application once per model on October 5, 2026, with the d1 Playground's comparison script. GPT-6.1 Sol and Claude Opus 5.5 get each request as one chat message and answer in JSON, at their default reasoning setting. Where many questions share one input, such as the filter's tickets or the folders' passages, the chat models answer them in batches. Costs use list prices, without prompt-cache discounts. d1's costs use $0.04 per million input tokens. Time is per run, with up to 8 requests in flight. A Smart Filter run is one query over 150 tickets. A Smart Folders run files 105 passages, and its cost is per 1,000 passages. The filter's quality is its F1 score against hand labels, which counts both missed and wrong matches. The other text applications count the goals, questions, passages or needed outputs handled correctly. We wrote six of the 15 code questions and two of the four compaction sessions after d1's pipeline was set. In Visual Inspection, every model sees a good part from the same line next to the part to inspect.