Liquid AI发布d1-3B与d1-omni-600M决策模型
Liquid AI Releases Open-Weight d1-3B and d1-omni-600M: Multimodal Decision Models With Zero Output Tokens
决策模型范式值得关注,d1-3B在同等参数下刷新了Benchmark记录且延迟极低,适合做Agent路由和审核的同学参考。
Liquid AI has released Open d1, two open-weight multimodal models in its d1 decision model family. d1-3B reads text and images. d1-omni-600M reads text with an image, or text with audio. Neither model writes text. Each returns calibrated, typed answers in one forward pass with zero output tokens. The target is real-time decisions on the NVIDIA stack: DGX servers, RTX workstations, and Jetson edge boards.
Liquid AI 发布了 Open d1,这是其 d1 决策模型系列中的两个开放权重多模态模型。d1-3B 可读取文本和图像。d1-omni-600M 可读取带图像的文本或带音频的文本。这两个模型均不生成文本。它们在一次前向传播中返回经过校准且类型明确的回答,输出 token 数为零。目标是在 NVIDIA 堆栈上实现实时决策:包括 DGX 服务器、RTX 工作站和 Jetson 边缘设备板卡。
Is it deployable? Yes. Both checkpoints are on Hugging Face, load through Transformers, and have day-one llama.cpp support. The LFM Open License v1.0 allows free commercial use below $10 million in annual revenue. d1-omni-600M is an early research release with no published latency figures.
是否可部署?可以。两个检查点(checkpoint)均已上传至 Hugging Face,可通过 Transformers 加载,并支持首日 llama.cpp 兼容。LFM Open License v1.0 允许年收入低于 1000 万美元的企业免费商用。d1-omni-600M 是一个早期研究版本,尚未公布延迟数据。
What is a decision model?
什么是决策模型?
A generative LLM writes its answer token by token, and your code parses it. A decision model takes a state and a set of named questions. It reads them once and returns a probability for every allowed answer. The Liquid AI define three question types:
生成式大语言模型(LLM)逐 token 生成答案,你的代码负责解析它。而决策模型接收一个状态和一组命名问题,一次性读取后为每个允许的答案返回概率。Liquid AI 定义了三种问题类型:
- noul: a yes or no question, returned as P(yes).
- choice: one label from named options, with a full probability distribution.
- score: a probability-weighted position on an ordered rubric of 2 to 10 levels.
- noul:是或否的问题,以 P(yes) 形式返回。
- choice:从命名选项中选择一个标签,并提供完整的概率分布。
- score:在 2 到 10 级的有序评分标准上,返回概率加权的得分位置。
Several questions can share one state in a single call. Every response reports output_tokens: 0. Liquid AI recommends d1 for routing, moderation, intent classification, reranking, LLM-as-a-judge scoring, agent guardrails, and visual inspection. Neither checkpoint is a chat model.
多个问题可以在单次调用中共享同一个状态。每个响应都报告 output_tokens: 0。Liquid AI 推荐将 d1 用于路由、内容审核、意图分类、重排序、LLM-as-a-judge 评分、智能体护栏(agent guardrails)以及视觉检查。这两个检查点均不是聊天模型。
How are d1-3B and d1-omni-600M built?
d1-3B 和 d1-omni-600M 是如何构建的?
d1-3B has 3.12B parameters and starts from LFM2.5-VL-3B, a decoder-only vision-language model. Liquid AI averaged the weights of LFM2.5-2.6B with that model’s text backbone. It then fine-tuned several checkpoints with different seeds and data mixtures, and merged them again. It uses a 400M SigLIP2 NaFlex vision encoder and a 32,768-token context. Long inputs, shuffled answer options, and fixing data shortcuts mattered more than advanced techniques.
d1-3B 拥有 31.2 亿参数,基于 LFM2.5-VL-3B 构建,这是一个仅解码器的视觉-语言模型。Liquid AI 将 LFM2.5-2.6B 的权重与该模型的文本主干进行了平均融合。随后,他们使用不同的随机种子和数据混合比例微调了多个检查点,并再次进行合并。它使用了 4 亿参数的 SigLIP2 NaFlex 视觉编码器,上下文长度为 32,768 个 token。长输入、打乱的答案选项以及修复数据捷径(data shortcuts)比高级技术更为重要。
d1-omni-600M has 587M parameters and starts from LFM2.5-Encoder-350M, a bidirectional encoder. That covers a 381M shared trunk and decision head, a 94M vision encoder, and a 112M audio encoder. The audio encoder is a 17-layer FastConformer. Context is 16,384 tokens. Audio clips are capped at 30 seconds. A request carries images or audio, never both. Audio training covered English speaker-to-assistant requests only.
d1-omni-600M 拥有 5.87 亿参数,基于 LFM2.5-Encoder-350M 构建,这是一个双向编码器。其中包括 3.81 亿的共享主干和决策头、9400 万的视觉编码器和 1.12 亿的音频编码器。音频编码器是一个 17 层的 FastConformer。上下文长度为 16,384 个 token。音频片段限制在 30 秒以内。请求中携带图像或音频,但不会同时携带两者。音频训练仅涵盖英语用户到助手的请求。
Where does Open d1 fit best?
Open d1 最适合应用于哪些场景?
- Support and ticket triage: One d1-3B call can answer a yes or no refund check, pick the owning team, and rate urgency over the same message, with zero output tokens to parse.
- Real-time visual inspection and moderation at the edge: d1-3B reads a 384px image in 35 ms on Jetson AGX Thor, and Liquid AI’s Open d1 Arcade runs it frame by frame on live camera input for content moderation and gesture control.
- Voice-command routing on small devices: d1-omni-600M takes up to 30 seconds of speech alongside text and returns the speaker’s intent or topic directly. Its card lists voice-command routing and agent guardrails among its recommended uses.
- 支持与工单分类:一次 d1-3B 调用即可回答是或否的退款检查、确定负责团队,并对同一消息进行紧急程度评级,且无需解析任何输出 token。
- 边缘端的实时视觉检查与审核:d1-3B 在 Jetson AGX Thor 上读取 384px 图像仅需 35 毫秒,Liquid AI 的 Open d1 Arcade 则在实时摄像头输入上逐帧运行该模型,用于内容审核和手势控制。
- 小型设备上的语音命令路由:d1-omni-600M 可接收长达 30 秒的语音及文本输入,并直接返回说话者的意图或主题。其卡片将语音命令路由和智能体护栏列为推荐用途之一。
How does d1 perform on benchmarks?
d1 在基准测试中的表现如何?
On Decision Index v0.2.1, d1-3B scores 48.57. That beats every model under 10B and edges Decider 35B-A3B (47.11). Only Winnow-12B scores higher at 50.02. Liquid AI ran the official scorer itself, so d1 scores are not leaderboard submissions. d1-3B leads the Tools (74.5) and Arts (36.3) categories but trails on Knowledge (23.8).
在 Decision Index v0.2.1 上,d1-3B 得分为 48.57。这超过了所有参数量低于 10B 的模型,并略高于 Decider 35B-A3B(47.11)。只有 Winnow-12B 得分更高,为 50.02。Liquid AI 自行运行了官方评分器,因此 d1 的分数并非排行榜提交结果。d1-3B 在 Tools(74.5)和 Arts(36.3)类别中领先,但在 Knowledge(23.8)类别中落后。
Across seven public text benchmarks, d1-3B averages 82.9, ahead of Decider 4B at 81.1. d1-omni-600M averages 78.4 and posts the top Civil Comments (95.8) and PAWS-X (79.5) scores. On 11 image benchmarks, d1-3B averages 74.1 against 73.9 for its base model. Liquid AI calls audio decision benchmarks an open problem.
在七个公开文本基准测试中,d1-3B 平均得分为 82.9,高于 Decider 4B 的 81.1。d1-omni-600M 平均得分为 78.4,并在 Civil Comments(95.8)和 PAWS-X(79.5)上取得最高分。在 11 个图像基准测试中,d1-3B 平均得分为 74.1,而其基础模型为 73.9。Liquid AI 称音频决策基准测试仍是一个开放性问题。
How fast is d1-3B on NVIDIA hardware?
d1-3B 在 NVIDIA 硬件上的速度如何?
Liquid AI measured end-to-end latency, one warm request at a time. One question takes 8 ms on an RTX 4090 and 9 ms on an AMD MI325X. On Jetson, AGX Thor takes 16 ms, AGX Orin takes 26 ms, and Orin Nano takes 50 ms. On Jetson AGX Thor, three questions over one state take 20 ms versus 16 ms for one. The RTX 4090 figure uses model.compile(mode="reduce-overhead"). Without it, one question takes 16 ms. NVIDIA’s Jetson AI Lab also hosts d1-3B guides.
Liquid AI 测量了端到端延迟,每次仅处理一个预热请求。在 RTX 4090 上,一个问题需 8 毫秒;在 AMD MI325X 上需 9 毫秒。在 Jetson 平台上,AGX Thor 需 16 毫秒,AGX Orin 需 26 毫秒,Orin Nano 需 50 毫秒。在 Jetson AGX Thor 上,跨越一个状态处理三个问题需 20 毫秒,而处理单个问题需 16 毫秒。RTX 4090 的数据使用了 model.compile(mode="reduce-overhead")。若不使用该配置,处理一个问题需 16 毫秒。NVIDIA 的 Jetson AI Lab 也提供了 d1-3B 的使用指南。
How does Open d1 compare with other open decision models?
Open d1 与其他开源决策模型相比如何?
| Feature | d1-3B | d1-omni-600M | Decider 4B | Decider 35B-A3B | Winnow-12B |
|---|---|---|---|---|---|
| Maker | Liquid AI | Liquid AI | Mapika (independent) | Mapika (independent) | EldanRing (independent) |
| Parameters | 3.12B | 587M | 4.2B | 34.7B total, 3B active | 12B |
| Base model | LFM2.5-VL-3B | LFM2.5-Encoder-350M | Qwen3.5-4B-Base | Qwen3.5-35B-A3B-Base | Gemma 4 12B IT |
| Inputs | Text, images | Text + image, or text + audio | Text | Text | Text, images |
| Context | 32,768 | 16,384 | 32K-token state | 32K-token state | 65,536 (Q8 tested) |
| Decision Index v0.2.1 | 48.57 | 15.95 | 40.70 | 47.11 | 50.02 |
| License | LFM Open v1.0 | LFM Open v1.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 |
| Published latency | 8 ms per question, RTX 4090 | Not published | 5.2 ms per 3-question request, B300 | 47 ms per 3-question request, B300 | 143 ms cached 4-question request near 64K context, RTX 5070 Ti |
| 特性 | d1-3B | d1-omni-600M | Decider 4B | Decider 35B-A3B | Winnow-12B |
|---|---|---|---|---|---|
| 制造商 | Liquid AI | Liquid AI | Mapika(独立) | Mapika(独立) | EldanRing(独立) |
| 参数量 | 3.12B | 587M | 4.2B | 总计 34.7B,激活 3B | 12B |
| 基础模型 | LFM2.5-VL-3B | LFM2.5-Encoder-350M | Qwen3.5-4B-Base | Qwen3.5-35B-A3B-Base | Gemma 4 12B IT |
| 输入 | 文本、图像 | 文本+图像,或文本+音频 | 文本 | 文本 | 文本、图像 |
| 上下文窗口 | 32,768 | 16,384 | 32K-token 状态 | 32K-token 状态 | 65,536(Q8 测试) |
| Decision Index v0.2.1 | 48.57 | 15.95 | 40.70 | 47.11 | 50.02 |
| 许可证 | LFM Open v1.0 | LFM Open v1.0 | Apache 2.0 | Apache 2.0 | Apache 2.0 |
| 发布延迟 | 每问题 8 ms,RTX 4090 | 未公布 | 每 3 个问题请求 5.2 ms,B300 | 每 3 个问题请求 47 ms,B300 | 近 64K 上下文的缓存 4 个问题请求 143 ms,RTX 5070 Ti |
Sources: d1-3B, d1-omni-600M, Decider 4B, Decider 35B-A3B, Winnow-12B model cards. Index scores as listed on the d1 cards. Latencies use different hardware and workloads, so they are not directly comparable.
来源:d1-3B、d1-omni-600M、Decider 4B、Decider 35B-A3B、Winnow-12B 模型卡片。索引分数如 d1 卡片所列。延迟使用不同的硬件和工作负载,因此不可直接比较。
How do you run d1 locally?
如何在本地运行 d1?
d1-3B needs transformers>=5.14 and trust_remote_code=True. d1-omni-600M needs transformers>=5.15. Both expose system_one(state, questions) for one state and system_one_batch for packed requests. Liquid AI recommends float16 for d1-omni-600M on GPU, since bfloat16 changed some top answers. You can try ten camera-driven d1-3B demos in the Open d1 Arcade space. With NVIDIA, Liquid AI also showed d1-3B navigating Isaac Sim from a Jetson.
d1-3B 需要 transformers>=5.14 和 trust_remote_code=True。d1-omni-600M 需要 transformers>=5.15。两者均暴露 system_one(state, questions) 用于单个状态,以及 system_one_batch 用于打包请求。Liquid AI 建议在 GPU 上使用 float16 运行 d1-omni-600M,因为 bfloat16 会改变部分最佳答案。你可以在 Open d1 Arcade 空间中尝试十个由摄像头驱动的 d1-3B 演示。借助 NVIDIA,Liquid AI 还展示了 d1-3B 在 Jetson 上导航 Isaac Sim。
Key Takeaways
关键要点
- d1 models output probabilities over fixed options in one pass, never generated tokens.
- d1-3B scores 48.57 on Decision Index v0.2.1, the best result under 10B.
- d1-3B answers one question in 8 ms on an RTX 4090.
- d1-omni-600M handles text with images or audio at 587M parameters.
- License is free for commercial use below $10M annual revenue.
- d1 模型在一次传递中对固定选项输出概率,不生成任何 token。
- d1-3B 在 Decision Index v0.2.1 上得分 48.57,是 10B 以下参数的最佳结果。
- d1-3B 在 RTX 4090 上回答一个问题仅需 8 ms。
- d1-omni-600M 以 587M 参数处理带图像或音频的文本。
- 许可证允许年收入低于 1000 万美元的商业用途免费使用。
Check out the d1-3B, d1-omni-600M and Technical details. All credit goes to the researcher of this project. Also, feel free to follow us on Twitter and don’t forget to join our 150k+ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.
查看 d1-3B、d1-omni-600M 和技术细节。所有功劳归于本项目的研究者。此外,欢迎在 Twitter 上关注我们,别忘了加入我们拥有 15 万+成员的 ML SubReddit 并订阅我们的新闻通讯。等等!你在 Telegram 上吗?现在你也可以加入我们的 Telegram 群组。
[Sponsored] The web is the one API most agents are missing. Databases, calendars and repos have APIs. The open web mostly doesn’t. The TinyFish MCP server gives any MCP client four tools: TinySearch, TinyFetch (full pages as markdown, JavaScript included), TinyBrowser for logins and forms, and TinyAgent for multi-step jobs. Search and Fetch are free.
[赞助内容] 网页是大多数智能体所缺失的 API。数据库、日历和代码仓库都有 API,而开放的网络大多没有。TinyFish MCP 服务器为任何 MCP 客户端提供四种工具:TinySearch(搜索)、TinyFetch(将完整页面转换为包含 JavaScript 的 Markdown)、TinyBrowser(用于登录和表单操作)以及 TinyAgent(用于多步骤任务)。搜索和抓取功能免费。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力