Liquid AIwhat is the fastest decision model you say?
16 7 1 129 11.2K 40
@ramin_m_h reposted
d1 from @liquidai is live on AI Gateway.
𝚕𝚒𝚚𝚞𝚒𝚍/𝚍𝟷
Make decisions with vision support to classify, route, and score.
vercel.com/changelog/liqu…
𝚕𝚒𝚚𝚞𝚒𝚍/𝚍𝟷
Make decisions with vision support to classify, route, and score.
vercel.com/changelog/liqu…
5 17 1 62 6.2K 6
@ramin_m_h reposted
Vercel Developers@vercel_dev · 12hd1 from @liquidai is live on AI Gateway.
𝚕𝚒𝚚𝚞𝚒𝚍/𝚍𝟷
Make decisions with vision support to classify, route, and score.
vercel.com/changelog/liqu…
Open quoted post →𝚕𝚒𝚚𝚞𝚒𝚍/𝚍𝟷
Make decisions with vision support to classify, route, and score.
vercel.com/changelog/liqu…
2 4 0 51 2.9K 2
@ramin_m_h reposted
Song 2 2 1 24 2.4K 2
@ramin_m_h reposted
now available as a HF space 🤗
i challenge @songdng and @SinoueG to beat my score.
(brb getting a new wardrobe)
huggingface.co/spaces/LiquidA…
i challenge @songdng and @SinoueG to beat my score.
(brb getting a new wardrobe)
huggingface.co/spaces/LiquidA…
clem 🤗 1 5 1 18 3.4K 0
so true! we are hiring for this, exactly this!!
adel 🌟@adelwu_ · Oct 8literal hardest role to hire for rn and every ai startup wants someone
> chronically online (knows trends)
> has taste & can create (new media)
> understands ai & technical
> can execute with quality (operator)
such a small niche of folks who can do this
Open quoted post →> chronically online (knows trends)
> has taste & can create (new media)
> understands ai & technical
> can execute with quality (operator)
such a small niche of folks who can do this
7 0 0 40 2.6K 14
@ramin_m_h reposted
Wow. Less than a month ago, Jev released and decision models took over X. 3 weeks later, @liquidai shipped a 3B that goes head to head with Jev -- on a Mac, 30 ms latency, and *with* vision...
The frontier is reaching local hardware faster than anyone could've anticipated.
The frontier is reaching local hardware faster than anyone could've anticipated.
Liquid AI@liquidai · Oct 7Today we release Open d1: two open-weight multimodal models in our d1 decision model family.
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
Open quoted post →
3 6 0 32 1.9K 11
@ramin_m_h reposted
AI models are getting small enough to make real-time decisions directly on your devices.
Liquid AI just released two open-weight decision models, and the performance is impressive.
Their 3B model can make a decision in just 8ms on an NVIDIA RTX 4090 and 50ms on a Jetson Orin Nano.
No need to send every request to the cloud.
Think robots, smart cameras, and AI agents that can understand what's happening and react in real time.
I think we're going to see a lot more AI running locally, especially as smaller, specialized models get this capable.
Liquid AI just released two open-weight decision models, and the performance is impressive.
Their 3B model can make a decision in just 8ms on an NVIDIA RTX 4090 and 50ms on a Jetson Orin Nano.
No need to send every request to the cloud.
Think robots, smart cameras, and AI agents that can understand what's happening and react in real time.
I think we're going to see a lot more AI running locally, especially as smaller, specialized models get this capable.
Liquid AI@liquidai · Oct 7Today we release Open d1: two open-weight multimodal models in our d1 decision model family.
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
Open quoted post →
13 9 1 37 5.2K 7
@ramin_m_h reposted
😱天呐！3B 本地决策模型把 OpenAI Decision 给彻底干报废了！
Atomic Chat 实测（单卡 RTX 3090 vs 云端 API）：
🟢 Liquid d1-3B（本地）
· 决策 8–20 ms
· 约 82 步/秒
· 吃下 86 个苹果
· 零输出 token，一次前向出概率
🔴 OpenAI Decision（云端）
· 单次调用 ~345 ms（视频里冲到 1.8 秒）
· 同样时间只吃到 3 个苹果
· 实时循环直接被网络往返拖死
3B 开源权重，边缘设备也能跑，云端大厂决策 API 在速度上被完虐。
本地决策模型，这才叫真实时！
#LiquidAI #d1 #OpenAI #LocalAI #EdgeAI #DecisionModels
x.com/nft_chen/statu… x.com/atomic_chat_hq…
Atomic Chat 实测（单卡 RTX 3090 vs 云端 API）：
🟢 Liquid d1-3B（本地）
· 决策 8–20 ms
· 约 82 步/秒
· 吃下 86 个苹果
· 零输出 token，一次前向出概率
🔴 OpenAI Decision（云端）
· 单次调用 ~345 ms（视频里冲到 1.8 秒）
· 同样时间只吃到 3 个苹果
· 实时循环直接被网络往返拖死
3B 开源权重，边缘设备也能跑，云端大厂决策 API 在速度上被完虐。
本地决策模型，这才叫真实时！
#LiquidAI #d1 #OpenAI #LocalAI #EdgeAI #DecisionModels
x.com/nft_chen/statu… x.com/atomic_chat_hq…
0:23
SuSu_酥酥👅@NFT_Chen · Oct 8🤯我靠！Liquid AI 把决策从大模型里单独拆出来了，本地开源能做到毫秒级延迟！在 4090 上 8 ms，Jetson 上也能实时跑，还带视觉！
Open d1 不写字，一次前向直接吐概率。分类、路由、打分、护栏，不用再让 LLM 生成 JSON 再解析。
性能亮点：
🔹 d1-3B 在 Decision Index v0.2.1 拿下 48.57，10B 以下第一，追平约 12 倍大的 Decider 35B-A3B
🔹单问题延迟：RTX 4090 8ms / Jetson AGX Thor 16ms / AGX Orin 26ms / Orin Nano 50ms
🔹文本七项基准均值 82.9，超过 Decider 4B
🔹d1-3B：文本+图像；实验版 d1-omni-600M：文本+图像或文本+音频
🔹零输出 token，首日 llama.cpp，DGX 到 Jetson 都能跑
上手只要四步：
1️⃣pip install “transformers>=5.14” torch torchvision pillow
2️⃣AutoModel.from_pretrained(“LiquidAI/d1-3B”, trust_remote_code=True)
3️⃣用 noul / choice / score 写问题
4️⃣model.system_one(state, questions) 一次出答案
权重： huggingface.co/LiquidAI/d1-3B huggingface.co/LiquidAI/d1-om… 博客：liquid.ai/blog/d1-open
#LiquidAI #OpenD1 #EdgeAI #HuggingFace #Jetson #DecisionModel #决策模型 #Jev
Open d1 不写字，一次前向直接吐概率。分类、路由、打分、护栏，不用再让 LLM 生成 JSON 再解析。
性能亮点：
🔹 d1-3B 在 Decision Index v0.2.1 拿下 48.57，10B 以下第一，追平约 12 倍大的 Decider 35B-A3B
🔹单问题延迟：RTX 4090 8ms / Jetson AGX Thor 16ms / AGX Orin 26ms / Orin Nano 50ms
🔹文本七项基准均值 82.9，超过 Decider 4B
🔹d1-3B：文本+图像；实验版 d1-omni-600M：文本+图像或文本+音频
🔹零输出 token，首日 llama.cpp，DGX 到 Jetson 都能跑
上手只要四步：
1️⃣pip install “transformers>=5.14” torch torchvision pillow
2️⃣AutoModel.from_pretrained(“LiquidAI/d1-3B”, trust_remote_code=True)
3️⃣用 noul / choice / score 写问题
4️⃣model.system_one(state, questions) 一次出答案
权重： huggingface.co/LiquidAI/d1-3B huggingface.co/LiquidAI/d1-om… 博客：liquid.ai/blog/d1-open
#LiquidAI #OpenD1 #EdgeAI #HuggingFace #Jetson #DecisionModel #决策模型 #Jev
Open quoted post →
10 18 1 151 23.2K 154
@ramin_m_h reposted
Can't get enough of d1? Here are two more reasons to be excited:
- d1-3B: text + vision
- d1-omni-600M: text + (image or audio)
running these models locally means going beyond seeing the world to interacting with it in real time (shoutout to @NVIDIARobotics).
wild.
- d1-3B: text + vision
- d1-omni-600M: text + (image or audio)
running these models locally means going beyond seeing the world to interacting with it in real time (shoutout to @NVIDIARobotics).
wild.
0:42Liquid AI@liquidai · Oct 7Today we release Open d1: two open-weight multimodal models in our d1 decision model family.
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
Open quoted post →
2 2 0 27 1.4K 4
@ramin_m_h reposted
ついに、Jevクラスタ < 狭ぁ
で話題だったLiquid AI d1の重みが公開!!
場合によっては、このまま普通に現場に導入できる可能性もある？FA屋さんは注目必須ッスわ。
di は Liquid AIが公開した 「読む・見る・聞く → 分類や判断をする」判断型モデル!!
Sol-6.1以上Opus5.5以下の判断力がついにお手元のPCや小型端末で動かせる～～！！
たとえば、こんな仕事をさせるモデルです。
- 問い合わせを読んで「返金／技術相談／配送」に振り分ける
- 写真を見て「部品に不良があるか」を判定する
- 音声を聞いて「ユーザーが何をしたいか」を分類する
- AIエージェントに与えた候補から、次の行動を選ぶ
- ロボットのティーチング、ピッキング
マルチモーダルなJev
モデルは2種類
d1-3B：テキスト＋画像
約31億パラメータ。判定精度を重視したモデル
d1-omni-600M：テキスト＋画像、またはテキスト＋音声
約6億パラメータ。小型化を重視した実験版
公式測定では、d1-3BはRTX 4090で単一質問に 8ms、384px画像を含む判定に 17ms。
精度も、構造化された判断を測る Decision Index v0.2.1で48.57、比較した10B未満のモデル中で最上位。
マルチモーダル判断モデルは普通に社会実装していく中で利便性がめっちゃ高いんで、バンバンこういう流れが続いていくのは本当に嬉しいっすね。
x.com/liquidai/statu…
で話題だったLiquid AI d1の重みが公開!!
場合によっては、このまま普通に現場に導入できる可能性もある？FA屋さんは注目必須ッスわ。
di は Liquid AIが公開した 「読む・見る・聞く → 分類や判断をする」判断型モデル!!
Sol-6.1以上Opus5.5以下の判断力がついにお手元のPCや小型端末で動かせる～～！！
たとえば、こんな仕事をさせるモデルです。
- 問い合わせを読んで「返金／技術相談／配送」に振り分ける
- 写真を見て「部品に不良があるか」を判定する
- 音声を聞いて「ユーザーが何をしたいか」を分類する
- AIエージェントに与えた候補から、次の行動を選ぶ
- ロボットのティーチング、ピッキング
マルチモーダルなJev
モデルは2種類
d1-3B：テキスト＋画像
約31億パラメータ。判定精度を重視したモデル
d1-omni-600M：テキスト＋画像、またはテキスト＋音声
約6億パラメータ。小型化を重視した実験版
公式測定では、d1-3BはRTX 4090で単一質問に 8ms、384px画像を含む判定に 17ms。
精度も、構造化された判断を測る Decision Index v0.2.1で48.57、比較した10B未満のモデル中で最上位。
マルチモーダル判断モデルは普通に社会実装していく中で利便性がめっちゃ高いんで、バンバンこういう流れが続いていくのは本当に嬉しいっすね。
x.com/liquidai/statu…
0:42 3 21 2 184 10.3K 128
@ramin_m_h reposted
Congrats to the @liquidai team, great job @ramin_m_h @xanamini @maximelabonne !
If you thought 100–500 ms was fast for a model call, d1-3B just reset the bar: 8 ms (!) per decision on a single RTX 4090.
Our current molecular property specialists are 4B models too, so we know how hard it is to fit this much into 3B.
If you thought 100–500 ms was fast for a model call, d1-3B just reset the bar: 8 ms (!) per decision on a single RTX 4090.
Our current molecular property specialists are 4B models too, so we know how hard it is to fit this much into 3B.
Liquid AI@liquidai · Oct 7Today we release Open d1: two open-weight multimodal models in our d1 decision model family.
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
Open quoted post →
0 1 0 10 344 1
@ramin_m_h reposted
LiquidAI/d1-3B でマリオクリアできた✌
こちらは一定の条件で画像入力ありにしていて、
・画像入力なしの区間：レイテンシ 70ms 程度
・画像入力ありの区間：レイテンシ 100ms 程度
とかなり優秀 ＠ RTX 5090
やはりローカル jev しか勝たんのか？
こちらは一定の条件で画像入力ありにしていて、
・画像入力なしの区間：レイテンシ 70ms 程度
・画像入力ありの区間：レイテンシ 100ms 程度
とかなり優秀 ＠ RTX 5090
やはりローカル jev しか勝たんのか？
0:37
金のニワトリ@gosrum · Oct 7気づいたら画像入力も可能な意思決定モデル d1 がオープンウェイトで公開されていた...！これも試さなくては
Decisions API は日本からだと通信の影響でレイテンシが大きくかつ安定性に課題があったけど、ローカル運用ならこのあたりの問題が解決できる
huggingface.co/LiquidAI/d1-3B
Open quoted post →Decisions API は日本からだと通信の影響でレイテンシが大きくかつ安定性に課題があったけど、ローカル運用ならこのあたりの問題が解決できる
huggingface.co/LiquidAI/d1-3B
1 26 5 169 24.2K 92
@ramin_m_h reposted
Watch me turn @altryne into a decision model believer
Alex Volkov@altryne · Oct 8Tomorrow 8:30am PT I'm going live with ThursdAI.
OpenAI dropped 722 math papers and mathematicians are split. Plus Claude Haiku 5.5, GPT-6 for everyone, a trillion-param Mistral, Reflection's Beam, and decision models everywhere.
@maximelabonne from @liquidai is joining at 10:15am PT to talk open d1.
youtube.com/watch?v=xFKnS3…
OpenAI dropped 722 math papers and mathematicians are split. Plus Claude Haiku 5.5, GPT-6 for everyone, a trillion-param Mistral, Reflection's Beam, and decision models everywhere.
@maximelabonne from @liquidai is joining at 10:15am PT to talk open d1.
youtube.com/watch?v=xFKnS3…
Open quoted post →
4 3 0 21 2.8K 3
@ramin_m_h reposted
8ms、やっぱ地元は近くて速いから良いね！！
atomic.chat@atomic_chat_hq · Oct 7Local Liquid d1 3B crushed OpenAI Decision at Snake 🐍
@liquidai d1 on a single RTX 3090 decides in 8 ms so its snake made 82 moves a second and ate 86 apples while OpenAI's API took 345 ms a call and ate 3
Run decision models -> atomic.chat
@liquidai d1 on a single RTX 3090 decides in 8 ms so its snake made 82 moves a second and ate 86 apples while OpenAI's API took 345 ms a call and ate 3
Run decision models -> atomic.chat
0:23 0 11 4 105 13.4K 44
@ramin_m_h reposted
d1-3Bとd1-omni-600MをiPhone 17 Proで動かしてみました！
スマホでも結構高速に動くし、これは色々使えそうかも！
スマホでも結構高速に動くし、これは色々使えそうかも！
Liquid AI@liquidai · Oct 7Today we release Open d1: two open-weight multimodal models in our d1 decision model family.
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
Open quoted post →
2 6 2 62 3.2K 20
@ramin_m_h reposted
Super cool new decision models from @LiquidAI! These are like Jev, which is an emerging new class of AI models that make decisions in a single pass rather than generating tokens one by one.
But LiquidAI Open d1 are both open-weight and multimodal decision models running locally, making decisions in milliseconds!
The implications of decision models for robotics, autonomous laboratories, medical devices, and physical AI are staggering and very exciting! In particular, I very much like that they work with @nvidia Jetson at the edge, very relevant for robotics.
This will be the bridge of AI to physical world, where every task series are decision making for the next action.
But LiquidAI Open d1 are both open-weight and multimodal decision models running locally, making decisions in milliseconds!
The implications of decision models for robotics, autonomous laboratories, medical devices, and physical AI are staggering and very exciting! In particular, I very much like that they work with @nvidia Jetson at the edge, very relevant for robotics.
This will be the bridge of AI to physical world, where every task series are decision making for the next action.
Liquid AI@liquidai · Oct 7Today we release Open d1: two open-weight multimodal models in our d1 decision model family.
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
Open quoted post →
6 7 0 81 10.3K 15
@ramin_m_h reposted
Liquid AI just released Open d1, 2 open-weight decision models that never generate a single token. d1-3B (text + vision) and d1-omni-600M (text + image, or text + audio) read a state plus typed questions and return calibrated probabilities in 1 forward pass.
The core idea is to replace "generate JSON, then parse it" with a direct readout. Every question is a noul (P(yes)), a choice (a distribution over named options) or a score (a position on a 2 to 10 level rubric). Output tokens: always 0.
d1-3B scores 48.57 on Decision Index v0.2.1. That is the best result under 10B, and above Decider 35B-A3B (47.11), a model 12x its size. The trade-off: it trails on Knowledge (23.8 vs 33.8 for Winnow-12B). It is a decision engine, not a chat model.
Latency per question: 8 ms on RTX 4090, 16 ms on Jetson AGX Thor, 50 ms on Jetson Orin Nano. Day-one llama.cpp support......
Full analysis: marktechpost.com/2026/10/07/liq…
d1-3B: huggingface.co/LiquidAI/d1-3B
d1-omni-600M: huggingface.co/LiquidAI/d1-om…
Docs: docs.liquid.ai/lfm/models/dec…
@liquidai @nvidia
The core idea is to replace "generate JSON, then parse it" with a direct readout. Every question is a noul (P(yes)), a choice (a distribution over named options) or a score (a position on a 2 to 10 level rubric). Output tokens: always 0.
d1-3B scores 48.57 on Decision Index v0.2.1. That is the best result under 10B, and above Decider 35B-A3B (47.11), a model 12x its size. The trade-off: it trails on Knowledge (23.8 vs 33.8 for Winnow-12B). It is a decision engine, not a chat model.
Latency per question: 8 ms on RTX 4090, 16 ms on Jetson AGX Thor, 50 ms on Jetson Orin Nano. Day-one llama.cpp support......
Full analysis: marktechpost.com/2026/10/07/liq…
d1-3B: huggingface.co/LiquidAI/d1-3B
d1-omni-600M: huggingface.co/LiquidAI/d1-om…
Docs: docs.liquid.ai/lfm/models/dec…
@liquidai @nvidia
2 17 1 47 3.1K 28
WILD. 2 days ago at SF tech week I said: at the frontier, there is no fundamental tech moat! no one has a superior tech or a new tech that is not achievable by others (this of course is a function of compute). the ultimate moat is fast & deep integration.
This means we are in the product-level competition. which is good. on the consumer side you see muse, instinct, grokbot, hark, and a 1000 more personal ai initiatives! everyone is doing the same thing! the best product wins.
on the enterprise side, the game is around neolabs/neoclouds selling inference and customization tokens in their market/sector, modality, and mode of choice.
to show you what I mean: decision models. Jev is a great idea, built over 2 years in stealth and launched closed. 2 weeks later we shipped d1, and today we open-sourced two d1 models, with recipes.
everything is gonna be open source, and everyone is gonna be on the same page.
This means we are in the product-level competition. which is good. on the consumer side you see muse, instinct, grokbot, hark, and a 1000 more personal ai initiatives! everyone is doing the same thing! the best product wins.
on the enterprise side, the game is around neolabs/neoclouds selling inference and customization tokens in their market/sector, modality, and mode of choice.
to show you what I mean: decision models. Jev is a great idea, built over 2 years in stealth and launched closed. 2 weeks later we shipped d1, and today we open-sourced two d1 models, with recipes.
everything is gonna be open source, and everyone is gonna be on the same page.
Liquid AI@liquidai · Oct 7Today we release Open d1: two open-weight multimodal models in our d1 decision model family.
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
> d1-3B: text + vision
> d1-omni-600M: text + image or text + audio
> Real-time decision making anywhere, from data centers such as @nvidia DGX to RTX workstations to Jetson at the edge.
1/
Open quoted post →
7 10 2 66 5.2K 23
@ramin_m_h reposted
I also stole @maximelabonne Melee demo to run it with the 3B. But here the game never waits for the model: it has to react in real time or get destroyed by the CPU.
60 fps, a decision every ~15 ms, and it can beat the level 7 Fox CPU. Never trained on Melee, it just reads the game state and picks a move 😎
60 fps, a decision every ~15 ms, and it can beat the level 7 Fox CPU. Never trained on Melee, it just reads the game state and picks a move 😎
2:21 1 1 1 13 734 3