Research Engineer

Liquid AI releases open d1 models for multimodal edge decisions. Liquid AI released d1-3B for text and images and an experimental 600M model that pairs text with images or audio. By Ryan Merket - Published Oct 7th, 2026, 12:32pm CT - 4 MIN READ - 4 SOURCES - VIA Hugging Face Newsroom Listen to article. Liquid AI, co-founded by CEO Ramin Hasani, released two open-weight models in its d1 decision family on October 7th. The Hugging Face release introduces d1-3B, which accepts text and images, and experimental d1-omni-600M, which accepts text paired with either images or audio. Developers can download and run the models themselves, or try demos in Liquid AI's System One Arcade. Hasani has pursued the idea of building models that work within the limits of real devices since Liquid AI's founding in 2023. Before co-founding Liquid AI, Hasani was a CSAIL postdoctoral associate and a principal AI and machine-learning scientist at Vanguard. His research includes decision-making in complex systems, relevant to models designed to return a choice or score instead of composing an answer in prose. Liquid AI introduced an API-hosted d1 model in a company post on October 5th and said it planned to release open weights for upcoming models. Two days later, the Hugging Face post offered d1-3B and d1-omni-600M as open-weight downloads. Liquid AI is making separate decision-model checkpoints available for local deployment alongside its API. A model built to return a decision. Generative models produce tokens, even for narrow tasks such as sorting a support ticket or deciding whether an image shows a defect. Liquid AI's decision models instead process a state in a single forward pass and return structured answers. The release describes yes-or-no decisions, label selection and rubric-based scores as supported outputs. That format could suit routing, inspection and classification tasks where software needs an answer it can act on, rather than a paragraph it must parse. The two models use different foundations. d1-3B is built on Liquid AI's LFM2.5-VL-3B vision-language model. d1-omni-600M uses the LFM2.5-Encoder-350M bidirectional encoder, with added vision and audio encoders. Liquid AI labels the smaller model experimental and says it remains under development; it did not report speed numbers for it. The new downloads let developers evaluate local deployment, while the October 5th release described an API-hosted model. Liquid AI's announcement says both new models are available on Hugging Face and includes code showing d1-3B loaded through Transformers. It also instructs users to enable trust_remote_code=True. The release calls the models open-weight but does not state a license. Both Hugging Face model cards list LFM 1.0; teams evaluating deployment should review those terms for their intended use. Company-reported benchmarks and latency. In its Hugging Face release, Liquid AI reports a Decision Index 0.2.1 score of 48.57 for d1-3B, above the listed 4B and 9B models and Decider 35B-A3B at 47.11. For a separate comparison across seven public datasets, Liquid AI reports average scores of 82.9 for d1-3B and 78.4 for d1-omni-600M. Those figures exceed the listed Decider 4B and 2B averages, respectively. The October 7th announcement's benchmark tables do not include vision or audio scores. The d1-3B model card separately reports an average score of 74.1 across 11 public image benchmarks, compared with 73.9 for its LFM2.5-VL-3B base model. Liquid AI says the Decision Index v0.3 vision split is private and that audio decision benchmarks remain an open problem. The latency figures are also company-reported. In the same release, Liquid AI reports that d1-3B answered one question in 16 milliseconds on a Jetson AGX Thor, 26 milliseconds on a Jetson AGX Orin and 50 milliseconds on a Jetson Orin Nano. On the Thor, the reported time for three questions was 20 milliseconds. These results are specific to the tested devices and workloads; they do not establish performance across other hardware or tasks. Liquid AI's d1 models extend its device strategy to structured decisions and let developers run models outside a hosted API. Whether their speed and quality hold up in production tasks remains to be tested, particularly for the experimental audio-capable model. Liquid AI said its $250 million Series A, announced in December 2024 and led by AMD, would support model development and edge and on-premise product readiness, according to its funding announcement. The d1 release applies that stated goal to compact models that return machine-readable decisions locally. A separate Hugging Face benchmark tested d1 on September 30th across 400 cases and 2,000 decisions, reporting 0.742 accuracy; it did not directly replicate Liquid AI's seven-dataset comparison. RuntimeWire, Inc. get virtually zero search traffic. Please consider sharing this story. Signal check Noise Signal Where does this story land? Drag the dial. The startup Native device foundation models Liquid AI builds foundation models designed to run natively on devices outside of data centers, optimized for low latency, privacy, and real-world hardware constraints. Their LFM family spans compact to larger models and modalities, including LFM2.5-1.2B-Thinking (on-device reasoning under 1GB), LFM2-8B-A1B (efficient on-device mixture-of-experts), LFM2-24B-A2B (scaled-up LFM2 architecture), LFM2-VL series (efficient vision-language, including a 3B and a 450M model for structured visual intelligence), LFM2-Audio (end-to-end audio foundation model), and LFM2-ColBERT-350M for embeddings. They showcase tool-calling agents on consumer hardware and emphasize local, fast, private applications like meeting summarization. The company offers developer tooling such as LEAP and Liquid Apollo to make building with on-device AI easier, has documented use of Meta's ExecuTorch for efficient deployment, and highlights support on laptops with AMD Ryzen power. Notable collaborations and announcements include a partnership with Mercedes-Benz to scale embedded in-car intelligence, a multi-year partnership with Shopify to bring sub-20ms foundation models to core commerce experiences, demonstrations with AMD and Robotec.ai for embedded autonomy, and a $250M raise to scale capable and efficient general-purpose AI. A public playground allows users to try Liquid models directly. d1-omni-600M. Hugging Face model repository LiquidAI/d1-omni-600M, listed under the image-text-to-text task. This listing describes a public repository, not an inference endpoint. d1-3B. LFM2.5-Encoder-350M. Hugging Face model repository LiquidAI/LFM2.5-Encoder-350M, listed under the fill-mask task. This listing describes a public repository, not an inference endpoint. Hugging Face model repository LiquidAI/d1-3B, listed under the image-text-to-text task. This listing describes a public repository, not an inference endpoint. Hugging Face model repository Mapika/decider-4b, listed under the text-classification task. This listing describes a public repository, not an inference endpoint.

Multi-harness RL improves a small model across four agent interfaces. Federico Ulfo October 2, 2026 Researchers from Hugging Face and Liquid AI have released an open framework for training small language models inside multiple coding-agent harnesses. An agent harness controls the tools, context and execution loop around a model; changing it can substantially change how well the same weights perform. The framework combines OpenEnv, Harbor and TRL. A capture proxy records the exact tokens and generation probabilities needed for reinforcement learning while Claude Code, Codex, OpenCode and Mini-SWE-Agent run without changes to their code. On held-out SmolDataEnvs data-analysis tasks, multi-harness training raised LFM2.5-2.6B's average pass@1 from 42.2% to 54.2%, with improvements across all four harnesses. It also reduced tool calls by 31% on tasks that both the trained model and the baseline solved. A small efficiency bonus rewarded correct solutions that used fewer calls. Training only in OpenCode reached 52.3% overall and performed best within OpenCode. The 1.9-point overall gap between the two training approaches was within evaluation noise. The clearer advantage of multi-harness training was its broader efficiency gains and stronger results under Claude Code and Codex. Supervised fine-tuning on successful rollouts produced smaller accuracy gains. The work offers a practical way to train open models for several agent interfaces at once. These results come from a small model on data-analysis tasks, so they do not establish the same gains on large-scale software-engineering benchmarks or harnesses absent from training.

Liquid AI releases d1: A decision model that returns calibrated probabilities with zero output tokens. Reading Time: 14 mins read Liquid AI has released d1, a decision model built for structured choices instead of text generation. You give it context and a set of typed questions. It returns calibrated probabilities across a fixed set of outcomes in a single call, with zero generated tokens. The target is the work many teams still send to general LLMs: classification, ticket routing, scoring, moderation, reranking and LLM-as-judge checks. Is it deployable? Yes, today, as a hosted API. d1 runs on the Liquid API under the model name d1:free. Liquid's model library lists it as API only and not trainable, so there are no GGUF, MLX or ONNX weights to self-host. You may also like. What is a decision model? A decision model evaluates a situation and returns a typed answer from options you define before the call. It does not write text. In every response, usage.output_tokens is 0. Liquid AI's migration guide gives a simple rule: if the answer is one of N known options, use a decision model. If the model must compose a new string, keep your LLM. The 3 primitives: Noul, Choice and Score. * Noul is a yes/no question that returns a probability between 0 and 1. In Liquid's example, 'Is this message a complaint?' returned 0.999. * Choice picks one option from a named set. It returns the top pick, the full distribution and a confidence value. A double-charge ticket scored 0.9997 on 'billing.' * Score rates input on an ordered rubric and returns a probability-weighted position. Levels are indexed from 0, so a 4-level urgency rubric spans 0 to 3. A production outage scored 2.9995. You can mix all 3 types in one request. The model evaluates every question against the same state in one round trip. How a d1 API call works. Each request has 3 parts: the model, the state (plain text or a JSON object) and the questions. Calls go to POST https://api.liquid.ai/decisions/v1/systemone. Keys come from console.liquid.ai and start with liquid_. The clients are TypeSafe AI's typesafe-sdk for Python and @typesafe-ai/sdk for TypeScript. from typesafe_sdk import TypeSafeClient, Noul client = TypeSafeClient(api_key=os.environ["LIQUID_API_KEY"], base_url="https://api.liquid.ai") result = client.system_one( model="d1:free", state="I have been waiting over three weeks for my order...", questions={"is_complaint": Noul( instructions="Is this message a complaint from the customer?")},) print(result.answers["is_complaint"].noul) # 0.999 Why move LLM classification calls to d1. The migration guide lists the concrete differences against an LLM with structured output: * No billed output tokens: An LLM bills output even for a one-word label. * Predictable latency: There is no decoding loop that grows with output length. * No schema errors: Answers always match the question type, so malformed JSON and retries go away. * Usable uncertainty: You get calibrated probabilities instead of a self-reported number. * Fewer round trips: 3 sequential classification calls become 1. Probabilities make thresholds practical. Liquid's moderation example blocks above 0.8, allows below 0.2 and sends the middle band to human review. Its routing example falls back to the most capable model tier when router confidence drops below 0.5. Liquid also says repeated evaluations of the same input are more consistent, which reduces verdict flips. Keep an LLM for summarization, drafting, multi-turn chat, code generation and complex multi-step reasoning. Demo: road decider. Liquid's road-decider cookbook is a pixel-art survival racer. d1 uses a Choice question to pick left, center or right on every decision tick, about 2 to 5 times per second depending on game speed. The app is vanilla JavaScript on Node.js 18+, with a Vite proxy that keeps the API key server-side. A "Jev vs d1" mode races d1 against TypeSafe's typesafe/jev-1.13 through OpenRouter. The most useful lesson is about state design. Per-lane summaries with distance to the first obstacle produced more confident decisions than a raw grid of the road. Interactive explainer: d1 step by step. '+k+"+obj[k]+' ';el.appendChild(d);});setTimeout(function{var f=el.querySelectorAll('.fill');keys.forEach(function(k,i){f[i].style.width=Math.max(obj[k]*100,0.6)+'%'});},60);post;} function pills(el,list,cb){el.innerHTML=";list.forEach(function(x,i){var b=document.createElement('span');b.className="pill"+(i?":' on');b.textContent=x.name;b.onclick=function{el.querySelectorAll('.pill').forEach(function(p){p.classList.remove('on')});b.classList.add('on');cb(x)};el.appendChild(b)});cb(list[0]);} /* tabs */ var runners=; R.querySelectorAll('.tab').forEach(function(t){t.onclick=function{R.querySelectorAll('.tab').forEach(function(x){x.classList.remove('on')});R.querySelectorAll('.pane').forEach(function(x){x.classList.remove('on')});t.classList.add('on');$(t.dataset.p).classList.add('on');if(runners[t.dataset.p])runners[t.dataset.p];setTimeout(post,50)}}); /* p1 race */ var bill={billing:0.9997,account:0.0002,returns:0.00005,shipping:0.00003,technical:0.00002}; var toks=['{"','department','":',' "','billing','"}'];var timer; function race{clearInterval(timer);$('llmTok').innerHTML=";$('llmCnt').textContent="0";$('llmRes').textContent="...";$('d1Res').textContent="...";$('d1Bars').innerHTML=";$('llmParse').classList.remove('run');var c=$('d1Core');c.classList.remove('run');void c.offsetWidth;c.classList.add('run'); setTimeout(function{bars($('d1Bars'),bill,Object.keys(bill));$('d1Res').textContent="billing";},450); var i=0;timer=setInterval(function{if(i 0.999'},{name:'Harmful content',p:0.99,q:'state: "You\'re an absolute idiot... I\'m going to find out where you work."\nnoul: "Does this user message contain harmful, threatening, or abusive content?"\n=> 0.99'}]; function noulUpd{var hi=+$('hi').value,lo=+$('lo').value;$('hiV').textContent=hi.toFixed(2);$('loV').textContent=lo.toFixed(2);$('wiV').textContent=np.toFixed(3); R.querySelector('.gauge').style.setProperty('background','linear-gradient(90deg,#22c55e 0%,#22c55e '+lo*100+'%,#f59e0b '+lo*100+'%,#f59e0b '+hi*100+'%,#ef4444 '+hi*100+'%)','important'); $('needle').style.left="calc("+np*100+'% - 1.5px)';$('needleV').textContent=np.toFixed(3);var a=$('noulAct'); if(np>hi){a.textContent="block";a.style.background='#fde2e2';a.style.color="#b42318"}else if(np'+i+";$('scTicks').innerHTML=t;$('scNeedle').style.left="-1.5px";$('scV').textContent="0"; setTimeout(function{$('scNeedle').style.left="calc("+(x.s/mx*100)+'% - 1.5px)';$('scV').textContent=x.s},80);bars($('scBars'),x.p,Object.keys(x.p));$('scRes').textContent=x.s;$('scConf').textContent=x.c; $('scAct').textContent=x.n===4?(x.s>=2.5?'page on-call':x.s>=1.5?'escalate':'standard queue'):'escalate (SLA 4h)';})} runners.p4=scInit; /* p5 road */ var L=['left','center','right'],ROWS=6,ob=,cur=1;ob['0-1']=1;ob['2-3']=1; function road{var h="";for(var r=ROWS-1;r>=0;r-){for(var l=0;l'+(ob[k]?'traffic':'row '+(r+1))+'