←
Liquid AI1,741 posts
Liquid AI avatar

Liquid AI

@liquidai

Build efficient general-purpose AI at every scale.

◷ Joined March 2023▧ 263 Photos and videos
Liquid AI avatar
Liquid AI@liquidai·Oct 27

Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence. > 6 model instances, across text, audio, vision, and omni modalities, generative and decision models. > 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window. > 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding > 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages. > 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language. > 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input. > fine-tune on 1 GPU. > a new tri-hybrid architecture > architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs. > 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters. > 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒 Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device. 1/ 🧵

2:17
Ben @ NextGen Software avatar
⟳ Liquid AI retweeted
Ben @ NextGen Software@BenFromNG·18h

Liquid AI shipped d1.2 and d1.2-ultra, decision models for text, images, and audio. Ultra is #1 on the Artificial Analysis Decision Index. What I like: you pay only for inputs, and it answers "not found" instead of guessing. That's what you want in prod.

Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities. > #1 on the @ArtificialAnlys Decision Index (75) > Read a value from the input (text, number or date) > Find every passage of a kind in a text (spans) > Locate things in a picture (box for one, boxes for every instance) > Estimate a number (number with a range) > Pick every option that applies (choice) 1/ 🧵

Post attachment
Mathias Lechner avatar
⟳ Liquid AI retweeted
Mathias Lechner@mlech26l·18h

Best API decision model: Liquid AI d1.2-ultra Most popular open decision model: Liquid AI d1-2B

Post attachment
JJ avatar
⟳ Liquid AI retweeted
JJ@JosephJacks_·Nov 17

🥇

Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities. > #1 on the @ArtificialAnlys Decision Index (75) > Read a value from the input (text, number or date) > Find every passage of a kind in a text (spans) > Locate things in a picture (box for one, boxes for every instance) > Estimate a number (number with a range) > Pick every option that applies (choice) 1/ 🧵

Post attachment
Alok avatar
⟳ Liquid AI retweeted
Alok@analogalok·Nov 17

99% of people skimming this missed the actual alpha buried in the release notes. It’s not even the #1 Decision Index score (75) or the audio support. Liquid quietly changed the billing math: Under d1-ultra, asking 10 questions on one doc + image billed the entire state 10 times. Under d1.2-ultra, you pay for the text/image ONCE, and only pay the tiny token delta per extra question. Do the math on a 2k token doc + 1024px image + 10 questions: • d1-ultra: ~35,700 tokens • d1.2-ultra: ~3,840 tokens Same $0.042/1M headline price. Literally 9.3x cheaper in production. Plus: - 0.8 calibration error (vs 7.2 on GPT-6 Luna’s Decisions API) - 218ms median latency (claimed, ill test it myself and share later) & 2x context (66k) - Native spans (PII redaction), bounding boxes, counts, & "not found" abstention in one forward pass If you’re still paying an autoregressive LLM to hallucinate JSON schemas for routing and doc extraction, you are lighting margin on fire. Are you still using frontier chat models as glorified classifiers or have you moved to decision models yet?

Post attachment

Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities. > #1 on the @ArtificialAnlys Decision Index (75) > Read a value from the input (text, number or date) > Find every passage of a kind in a text (spans) > Locate things in a picture (box for one, boxes for every instance) > Estimate a number (number with a range) > Pick every option that applies (choice) 1/ 🧵

Post attachment
Moham avatar
⟳ Liquid AI retweeted
Moham@Moham_AI·Nov 17

If you’re still using Jev as your decision model in projects, it’s time to switch. d1.2-ultra is smarter, faster, and currently #1 on the Decision Index. Better calibration, new capabilities, and stronger overall performance.

Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities. > #1 on the @ArtificialAnlys Decision Index (75) > Read a value from the input (text, number or date) > Find every passage of a kind in a text (spans) > Locate things in a picture (box for one, boxes for every instance) > Estimate a number (number with a range) > Pick every option that applies (choice) 1/ 🧵

Post attachment
David Hendrickson avatar
⟳ Liquid AI retweeted
David Hendrickson@TeksEdge·Nov 16

🤔 We have a new Decision Model leader. 📊 Last week @NaceAI's Drex Max v1.1 topped the decision leaderboard. Now @liquidai just took the lead with d1.2-ultra. 👀But look at what these new decision models can actually do. Artificial Analysis Decision Index v2.1 ... 🥇 Liquid d1.2-ultra → 75 🧠 Drex Max v1.1 → 73 🧠 SubQ Snap 1.2 → 73 🧠 GLiDE v1.2 → 73 🧠 Liquid d1-ultra → 71 🧠 Liquid d1.2 → 69 Where's Jev? These aren't just models that choose which tool your AI agent should call but now they now process text, images AND audio and can ... 📄 Extract a number, date or value from a document 🔎 Find every passage of a particular type in a document 🖼️ Locate objects in an image with bounding boxes 🔢 Estimate a quantity with an uncertainty range 🎯 Select every answer that applies Ask it for the total on an invoice. It returns the value and where it found it. If the value isn't there? It can return "not found" instead of inventing a number. This is Liquid's reported accuracy ... 🧾 Receipt value extraction → 98% 🖼️ Object localization → 94% 🔢 Object counting (within one) → 93% These are internal d1.2-ultra evaluations. Pricing? 💰 d1.2 → $0.020/M input tokens 💰 d1.2-ultra → $0.042/M input tokens No output-token charges. .... AND here's the best part. Liquid already has downloadable decision models! 🔓 d1-2B 🔓 d1-Omni-2B 🔓 d1-3B 🔓 d1-omni-600M And their earlier d1-3B model answers a question quickly ... ⚡ RTX 4090 → 8ms ⚡ Jetson AGX Thor → 16ms ⚡ Jetson Orin Nano → 50ms Application. Use a small specialist running locally that handles your agent's routing, tool selection, document extraction and verification. The new d1.2 and d1.2-ultra are API models today. No downloadable weights or local hardware requirements announced. So @liquidai when do Localmaxxers get the open weights? 🔥 Links in reply.

Post attachment
Song avatar
⟳ Liquid AI retweeted
Song@songdng·Nov 16

Post attachment

Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities. > #1 on the @ArtificialAnlys Decision Index (75) > Read a value from the input (text, number or date) > Find every passage of a kind in a text (spans) > Locate things in a picture (box for one, boxes for every instance) > Estimate a number (number with a range) > Pick every option that applies (choice) 1/ 🧵

Post attachment
Ivan Fioravanti avatar
⟳ Liquid AI retweeted
Ivan Fioravanti@ivanfioravanti·Nov 16

d1-2B by @liquidai :Same model, three engines on a MacBook Pro M5 Max - MLX: 66 ms, 153 W - Core AI GPU: 67 ms, 120 W - Core AI Neural Engine: 197 ms, 37 W The Neural Engine is 3x slower (on M5 Max, I bet this is different on M4) and draws a quarter of the power. Same answers on all three. I had to use a video instead of live camera feed because ANE usage goes up as soon as you start the camera. Core AI export from the original weights, plus live GPU / Neural Engine / power charts: github.com/ivanfioravanti/ml… Download and have fun! Share your stats on other Apple Silicon machines! I'm now creating an iPhone version 😎

0:48
Justin Li 李泽天 avatar
⟳ Liquid AI retweeted
Justin Li 李泽天@justinli9527·Nov 16

No.1 Decision model with omnimodal

Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities. > #1 on the @ArtificialAnlys Decision Index (75) > Read a value from the input (text, number or date) > Find every passage of a kind in a text (spans) > Locate things in a picture (box for one, boxes for every instance) > Estimate a number (number with a range) > Pick every option that applies (choice) 1/ 🧵

Post attachment
Vasko avatar
⟳ Liquid AI retweeted
Vasko@RoliumGens·Nov 16

Introducing : Voltron-1 Mini. A small, compact coding agent. Voltron has been post-trained for coding and mathematics knowledge, using @liquidai's LFM 3 1.7B as a base. Voltron delivers better performances across several like benchmarks like livecodebench ( substract of it ), AIME 2026, HumanEval and a few other benchmarks. This is the first of many models I'm going to do in this class. I've been working very hard on my RSI pipelines, and have managed to build something very promising which I will use to improve my models. This means more, better checkpoints are coming. Also, this model was trained 80% of the time using a new, more efficient LorA algorithm, research paper dropping soon, people with consumer GPUs are gonna love it!!! 🤗HF link : huggingface.co/Akahsizrr/vol… ( next release is the inference engine guys btw ).

leonie avatar
⟳ Liquid AI retweeted
leonie@helloiamleonie·Nov 16

time to update your decision model.

Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities. > #1 on the @ArtificialAnlys Decision Index (75) > Read a value from the input (text, number or date) > Find every passage of a kind in a text (spans) > Locate things in a picture (box for one, boxes for every instance) > Estimate a number (number with a range) > Pick every option that applies (choice) 1/ 🧵

Post attachment
Maxime Labonne avatar
⟳ Liquid AI retweeted
Maxime Labonne@maximelabonne·Nov 16

Super proud of the traction d1 got since its release. Now with audio and many new primitives, we're ready to shape the next stage of decision models 🚀

Today, we release d1.2 and d1.2-ultra, the state-of-the-art omnimodal decision models. d1.2 and d1.2-ultra both can process text, images, and audio with new output capabilities. > #1 on the @ArtificialAnlys Decision Index (75) > Read a value from the input (text, number or date) > Find every passage of a kind in a text (spans) > Locate things in a picture (box for one, boxes for every instance) > Estimate a number (number with a range) > Pick every option that applies (choice) 1/ 🧵

Post attachment