Liquid AI 发布 d1.2 系列全模态决策模型

Liquid AI 刚刚发布了 d1.2 和 d1.2-ultra。决策模型不生成文本,而是在一组固定答案上返回经过校准的概率。Liquid AI d1.2 决策模型系列现在接受文本、图像和音频,还新增了 5 种答案类型。d1.2-ultra 在 Artificial Analysis Decision Index v2.1 上得分为 75,是该指数上决策模型中的最高分。你可以通过我们的 GPULeaderboard.com Chrome 扩展在此查看部署选项和规格。
d1.2 是 Liquid AI 提供的托管式决策模型,可在 1 次调用中回答关于文本、图像或音频的类型化问题。它为每个选项返回一个概率,并生成零个输出 Token,因此仅对输入计费。
d1.2-ultra 面向更困难的决策。d1.2 更快,成本不到其一半。模型 ID 为 d1.2(默认)和 d1.2-ultra。较早的 d1、d1.1 和 d1-ultra 现已弃用,并且不支持音频或新任务。
d1.2 在相同的端点和请求格式上新增了 5 种类型化答案:
音频可以作为问题的主题,例如说的是哪种语言,或是否能听到警报声。问题也可以通过语音提出。API 接受 16 位 PCM WAV,并按每秒音频 10 个 Token 计费。在留出的 ESC-50 和 FLEURS(43 种语言)录音上,两个模型正确回答了 95% 到 97% 的问题。
该指数综合了 6 项评估。AA-Paperclip 占 30% 的权重。EvalSafe-O*NET-AA、AA-Omniscience-Decision v1.1、HLE-MultiChoice v1.1 和 AA-Symmetry v1.1 各占 15%。When2Call 占 10%。其 Decision Score 会惩罚高置信度的错误答案。
在 Liquid AI 分享的 AA 图表上,d1.2-ultra 得分 75。Drex Max v1.1、SubQ Snap 1.2 和 GLiDE v1.2 以 73 分紧随其后。d1.2 得分 69,比已弃用的 d1-ultra 的 71 分低 2 分。Liquid AI 称,没有哪个模型能在更低价格下得分高于 d1.2。
一个经过校准的模型说 80% 时,大约有 80% 的情况下是对的,因此团队可以把低置信度的案例转交人工处理。在 13 项基准测试和 22,505 个问题中,Liquid AI 报告了以下校准误差:
注:这些数据来自 Liquid AI 自己的测试。
Liquid AI 使用的是公开数据集的留出切分;目前尚不存在公开基准测试。
在多标签选择上,d1.2 以 90% 对 87% 胜过 d1.2-ultra。
定价为每 100 万输入 Token 0.02 美元(d1.2)和 0.042 美元(d1.2-ultra)。新组织在 Liquid Console 上可获得 2 美元免费额度。
图像按每个 32×32 图块 1.5 个 Token 计费,因此一张 1024×1024 的图像为 1,536 个 Token。共享状态和媒体只计费一次;每增加一个问题,只增加该问题自身的 Token。
每次调用的请求限制:
端点为 https://api.liquid.ai/decisions/v1/systemone。
欢迎查看发布博客、d1 API 文档以及 AA Decision Index。你可以通过我们的 GPULeaderboard.com Chrome 扩展在这里查看部署选项和规格。
需要我们合作推广你的 GitHub 仓库、Hugging Face 页面、产品发布或网络研讨会等?欢迎与我们联系
文章《Liquid AI 发布 d1.2 和 d1.2-ultra:登顶 AA Decision Index 的全模态决策模型》最先出现在 MarkTechPost 上。
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Liquid AI Releases d1.2 and d1.2-ultra: Omnimodal Decision Models That Top the AA Decision Index)
Liquid AI just released d1.2 and d1.2-ultra. A decision model does not write text. It returns calibrated probabilities over a fixed set of answers. The Liquid AI d1.2 decision model family now accepts text, images and audio. It also adds 5 new answer types. d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1, the top score among decision models on the index. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.
TL;DR
- Size: Parameters not disclosed. Up to 65,536 input tokens per question, 128 questions per call.
- Runs on: Hosted Liquid API only. No weights yet; Liquid AI says it will open them “at some point”.
- Performance: d1.2-ultra 75 and d1.2 69 on the AA Decision Index v2.1.
- Best: 0.8-point calibration error for d1.2-ultra across 13 benchmarks (Liquid AI test).
What is Liquid AI d1.2?
d1.2 is a hosted decision model from Liquid AI that answers typed questions about text, images or audio in 1 call. It returns a probability per option and generates zero output tokens, so only input is billed.
d1.2-ultra targets harder decisions. d1.2 is faster and costs less than half as much. The model IDs are d1.2 (default) and d1.2-ultra. Earlier d1, d1.1 and d1-ultra are now deprecated and do not support audio or the new tasks.
What are the 5 new tasks in d1.2?
d1.2 adds 5 typed answers on the same endpoint and request format:
- Read a value (
text,number,date): returns the value plus the characters it came from. It answers “not found” instead of guessing. - Find every span (
spans): returns each matching passage, such as names or emails, with its position. - Locate objects (
box,boxes): returns 1 rectangle, or 1 per instance with a count. - Estimate a number (
numberwithrange): returns a best guess and a range, for example 12 within 10 to 14. - Pick every option that applies (
choicewithmultiple): returns each matching label with its own probability.
How does audio input work?
Audio can be the subject of a question, such as which language is spoken or whether a siren is audible. The question can also be spoken. The API takes 16-bit PCM WAV and counts 10 tokens per second of audio. On held-out ESC-50 and FLEURS (43 languages) recordings, both models answer 95% to 97% of questions correctly.
How does d1.2 score on the Artificial Analysis Decision Index?
The index combines 6 evaluations. AA-Paperclip carries 30% of the weight. EvalSafe-O*NET-AA, AA-Omniscience-Decision v1.1, HLE-MultiChoice v1.1 and AA-Symmetry v1.1 carry 15% each. When2Call carries 10%. Its Decision Score penalizes confident wrong answers.
On the AA chart shared by Liquid AI, d1.2-ultra scores 75. Drex Max v1.1, SubQ Snap 1.2 and GLiDE v1.2 follow at 73. d1.2 scores 69. That is 2 points below the deprecated d1-ultra at 71. Liquid AI argues no model scores higher than d1.2 for less money.
How well calibrated are the probabilities?
A calibrated model that says 80% is right about 80% of the time, so teams can route low-confidence cases to humans. Across 13 benchmarks and 22,505 questions, Liquid AI reports these calibration errors:
- d1.2-ultra: 0.8 points
- d1.2: 1.2
- Jev 1.14: 2.7
- Perplexity Decider v1.3: 3.7
- GPT-6 Luna (Decisions API): 7.2
Note: These figures come from Liquid AI’s own test.
How accurate are the new tasks?
Liquid AI used held-out splits of public datasets; no public benchmark exists yet.
| Task (dataset) | d1.2-ultra | d1.2 |
|---|---|---|
| Read a value from receipts (CORD, SROIE) | 98% | 96% |
| Find personal data, precision / recall (Nemotron PII) | 94% / 93% | 92% / 88% |
| Locate an object (RefCOCO, RefCOCO+, RefCOCOg) | 94% | 94% |
| Count objects within 1 (COCO) | 93% | 92% |
| Pick every object present (COCO) | 87% | 90% |
d1.2 beats d1.2-ultra on multi-label picks, 90% to 87%.
d1.2 vs other decision models
| Metrics | d1.2-ultra | d1.2 | Drex Max v1.1 (Nace.AI) | Decider v1.3 (Perplexity) | Jev 1.14 (TypeSafe) |
|---|---|---|---|---|---|
| AA Decision Index v2.1 | 75 | 69 | 73 | 70 | 66 |
| Calibration error (Liquid test) | 0.8 | 1.2 | Not reported | 3.7 | 2.7 |
| Input price per 1M tokens | $0.042 | $0.02 | $0.03 | $0.02 | $0.042 |
| Output price | $0 | $0 | $0 | $0 | $0 |
| Input limit | 65,536 per question | 65,536 per question | 262,144 | 262,144 | 65.5K |
| Inputs | Text, image, audio | Text, image, audio | Text, image | Text, image | Text |
| Parameters | Not disclosed | Not disclosed | Not disclosed | 27B base (Qwen3.8-27B) | Not disclosed |
| Weights | Planned, no date | Planned, no date | v1.1 not released | Apache 2.0 | Proprietary |
</button><div class="stxt" id="stxt">Press Next to step through a d1.2 request.</div></div>
<div class="note">Limits from Liquid AI's d1 docs: 65,536 input tokens per question, 4.5 MB request body. Parameter count and architecture are not disclosed.</div>
</div>
<div class="pane" data-p="1">
<div class="grid5">
<div class="tcard on" data-k="0"><b>Read a value</b><code>text · number · date</code></div>
<div class="tcard" data-k="1"><b>Find every span</b><code>spans</code></div>
<div class="tcard" data-k="2"><b>Locate objects</b><code>box · boxes</code></div>
<div class="tcard" data-k="3"><b>Estimate a number</b><code>number + range</code></div>
<div class="tcard" data-k="4"><b>Pick all that apply</b><code>choice + multiple</code></div>
</div>
<div class="demo"><div><div class="lbl">You ask</div><pre id="dq"></pre></div><div><div class="lbl">Shape of the answer (illustrative)</div><pre id="da"></pre></div></div>
<div class="acc" id="dacc"></div>
</div>
<div class="pane" data-p="2">
<div class="lbl">Artificial Analysis Decision Index v2.1 (higher is better)</div>
<div class="bars" id="dib"></div>
<div class="note">Scores from the AA chart shared by Liquid AI on Nov 16, 2026. Selected models shown. Index weights: AA-Paperclip 30%, When2Call 10%, 4 others 15% each.</div>
</div>
<div class="pane" data-p="3">
<div class="lbl">Calibration error, points (lower is better)</div>
<div class="bars" id="cab"></div>
<div class="note">Liquid AI test: 13 benchmarks, 22,505 questions answered by all 5 models. Not an independent evaluation. A well-calibrated model saying 80% is right about 80% of the time, so you can route low-confidence cases to human review.</div>
</div>
<div class="pane" data-p="4">
<div class="calc">
<div>
<label>Text state tokens: <b id="vt">1,000</b><input type="range" id="it" min="0" max="60000" step="500" value="1000"></label>
<label>1024x1024 images: <b id="vi">1</b> <small>(1,536 tokens each)</small><input type="range" id="ii" min="0" max="8" value="1"></label>
<label>Audio seconds: <b id="va">0</b> <small>(10 tokens per second)</small><input type="range" id="ia" min="0" max="300" step="5" value="0"></label>
<label>Questions: <b id="vq">3</b> × tokens each: <b id="vqt">50</b> <small>(your estimate)</small><input type="range" id="iq" min="1" max="128" value="3"><input type="range" id="iqt" min="10" max="500" step="10" value="50"></label>
</div>
<div class="out">
<div class="lbl" style="color:#D4C0FF!important">Cost per 1M requests</div>
<div class="row"><span>Tokens per request</span><b id="tok">0</b></div>
<div class="row"><span>d1.2 ($0.02 / 1M)</span><b id="c1">$0</b></div>
<div class="row"><span>d1.2-ultra ($0.042 / 1M)</span><b id="c2">$0</b></div>
<div class="big" style="margin-top:10px">d1.2 saves <span id="sv">0%</span></div>
<div class="note" style="color:#D4C0FF!important">Output tokens are always 0. State and media are billed once per request; each question adds its own tokens.</div>
</div>
</div>
</div>
<div class="ft"><span>Sources: <a href="https://www.liquid.ai/blog/d1-2" target="_blank" rel="noopener">liquid.ai/blog/d1-2</a> · <a href="https://docs.liquid.ai/lfm/models/d1" target="_blank" rel="noopener">docs.liquid.ai</a> · self-reported unless noted</span><b>© Marktechpost</b></div>
<script>
(function(){
var R=document.getElementById('mtp-d12x');
function post(){try{parent.postMessage({mtpH:R.offsetHeight+40,id:'mtp-d12x'},'*')}catch(e){}}
var tabs=R.querySelectorAll('.tab'),panes=R.querySelectorAll('.pane');
tabs.forEach(function(t){t.onclick=function(){tabs.forEach(function(x){x.classList.remove('on')});panes.forEach(function(x){x.classList.remove('on')});t.classList.add('on');var p=R.querySelector('[data-p="'+t.dataset.t+'"]');p.classList.add('on');if(t.dataset.t==='2')bars('dib',DI,80);if(t.dataset.t==='3')bars('cab',CA,8);setTimeout(post,60)}});
var S=['Press Next to step through a d1.2 request.','1. You send a state (text, JSON, images or audio) plus typed questions.','2. d1.2 reads the state and every question in a single forward pass.','3. It returns a probability per option and typed answers. No text is generated.','4. You set a threshold: confident answers act automatically, low-confidence ones go to a human.'];
var s=0,core=R.querySelector('#core');
function step(){R.querySelector('#stxt').textContent=S[s];R.querySelectorAll('[data-s="1"]').forEach(function(c){c.classList.toggle('lit',s>=1)});R.querySelector('#a1').classList.toggle('go',s>=2);R.querySelector('#a2').classList.toggle('go',s>=3);R.querySelectorAll('[data-s="3"]').forEach(function(c){c.classList.toggle('lit',s>=3)});if(s===2){core.classList.remove('pulse');void core.offsetWidth;core.classList.add('pulse')}}
R.querySelector('#next').onclick=function(){s=s<4?s+1:0;step()};R.querySelector('#prev').onclick=function(){s=s>0?s-1:0;step()};
var T=[
['state: invoice.png\nquestion: "How much is due?"\ntype: number','value: 1234.50\nsource: "1 234,50"\nfound: 0.99\n(if absent: "not found")','Receipts (CORD, SROIE): d1.2-ultra 98%, d1.2 96%'],
['state: support note (text)\nquestion: "Every piece of personal data"\ntype: spans','spans:\n name [start,end]\n phone [start,end]\n email [start,end]','PII precision / recall (Nemotron PII): d1.2-ultra 94% / 93%, d1.2 92% / 88%'],
['state: photo.jpg (exactly 1 image)\nquestion: "Every tree"\ntype: boxes','boxes: [x1,y1,x2,y2] ...\ncount: n\n(box: 1 rectangle or "not found")','Locate an object (RefCOCO family): 94% for both models'],
['state: crowd.jpg\nquestion: "How many people?"\ntype: number + range','estimate: 12\nrange: 10 to 14','Count within 1 (COCO): d1.2-ultra 93%, d1.2 92%'],
['state: support message\nquestion: "Which topics apply?"\ntype: choice, multiple: true','billing: p\ndelivery: p\n(each with its own probability)','Pick every object present (COCO): d1.2-ultra 87%, d1.2 90%']];
var tc=R.querySelectorAll('.tcard');
function demo(k){tc.forEach(function(c){c.classList.toggle('on',c.dataset.k==k)});R.querySelector('#dq').textContent=T[k][0];R.querySelector('#da').textContent=T[k][1];R.querySelector('#dacc').innerHTML='<b>Liquid AI internal test:</b> '+T[k][2]}
tc.forEach(function(c){c.onclick=function(){demo(+c.dataset.k)}});demo(0);
var DI=[['d1.2-ultra',75,'me'],['Drex Max v1.1',73],['SubQ Snap 1.2',73],['GLiDE v1.2',73],['Decider v1.3',70],['d1.2',69,'me2'],['Jev 1.14',66],['GPT-6 Luna (Decisions)',60]];
var CA=[['d1.2-ultra',0.8,'me'],['d1.2',1.2,'me2'],['Jev 1.14',2.7],['Perplexity Decider v1.3',3.7],['GPT-6 Luna (Decisions API)',7.2]];
function bars(id,d,mx){var el=R.querySelector('#'+id);el.innerHTML=d.map(function(r){return '<div class="bar"><span>'+r[0]+'</span><div class="track"><div class="fill '+(r[2]||'')+'" data-w="'+(r[1]/mx*100)+'"></div></div><span class="v">'+r[1]+'</span></div>'}).join('');setTimeout(function(){el.querySelectorAll('.fill').forEach(function(f){f.style.width=f.dataset.w+'%'})},40)}
function f(n){return n.toLocaleString('en-US')}
function calc(){var t=+R.querySelector('#it').value,i=+R.querySelector('#ii').value,a=+R.querySelector('#ia').value,q=+R.querySelector('#iq').value,qt=+R.querySelector('#iqt').value;
R.querySelector('#vt').textContent=f(t);R.querySelector('#vi').textContent=i;R.querySelector('#va').textContent=a;R.querySelector('#vq').textContent=q;R.querySelector('#vqt').textContent=qt;
var tok=t+i*1536+Math.ceil(a*10)+q*qt;R.querySelector('#tok').textContent=f(tok);
R.querySelector('#c1').textContent='$'+(tok*0.02).toFixed(2);R.querySelector('#c2').textContent='$'+(tok*0.042).toFixed(2);R.querySelector('#sv').textContent=Math.round((1-0.02/0.042)*100)+'%'}
R.querySelectorAll('input').forEach(function(x){x.oninput=calc});calc();
window.addEventListener('load',post);window.addEventListener('resize',post);setTimeout(post,300);
})();
</script>
</div>
</body></html>">
Pricing and deployment
Pricing is $0.02 per 1M input tokens for d1.2 and $0.042 for d1.2-ultra. New organizations get $2 in free credit on Liquid Console.
Images cost 1.5 tokens per 32×32 patch, so a 1024×1024 image is 1,536 tokens. Shared state and media are billed once; each extra question adds only its own tokens.
Request limits per call:
- 128 questions and a 4.5 MB request body
- 8 images, or exactly 1 for
boxandboxes - 8 audio clips and 300 seconds of total audio
The endpoint is https://api.liquid.ai/decisions/v1/systemone.
Key Takeaways
- d1.2-ultra scores 75 on the AA Decision Index v2.1, first among decision models.
- Both models now read text, images and audio with 0 output tokens.
- 5 new answer types: values, spans, boxes, ranged estimates, multi-label choices.
- d1.2-ultra reports 0.8-point calibration error vs 3.7 for Perplexity Decider v1.3.
- Hosted only for now: no weights, parameter counts or latency figures.
Check out the launch blog, the d1 API docs and the AA Decision Index. You can check the deployment options and specs here through our GPULeaderboard.com chrome extension.
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.
Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us
The post Liquid AI Releases d1.2 and d1.2-ultra: Omnimodal Decision Models That Top the AA Decision Index appeared first on MarkTechPost.
正文保留英文原文,标题与摘要由机器翻译。
原文发布于 MarkTechPost(Liquid AI Releases d1.2 and d1.2-ultra: Omnimodal Decision Models That Top the AA Decision Index),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。