Liquid AI

We love seeing what developers build with LFMs. Developer Lakshya Gupta felt journals are personal and should stay that way, so he built SelfLink, a journaling app powered by LFM2-2.6B. The model runs entirely on-device, reading your entries to surface patterns and insights, with nothing ever leaving your device. Try it: realityplay.io/selflink

We're celebrating an exciting milestone in our partnership with Shopify: our LFMs have now processed 1 billion requests on Shopify’s platform! Read more about our multi-year partnership here: https://lnkd.in/eeKMAQcg

We're celebrating an exciting milestone in our partnership with Shopify: our LFMs have now processed 1 billion requests on Shopify’s platform! Read more about our multi-year partnership here: https://lnkd.in/eeKMAQcg

We are at [ICML] Int'l Conference on Machine Learning 2026 in Seoul. Booth B116. We're hiring globally, including post-training and applied ML roles in our Tokyo office. Stop by our booth to talk! 私たちはICML 2026（ソウル）のブースB116に出展しています。 世界中で採用を行っており、東京オフィスではポストトレーニングおよび応用機械学習（Applied ML）のポジションも募集しています。ぜひブースにお立ち寄りいただき、お気軽にお話ししましょう！

Today, we release Antidoom, an open-source method that removes a common failure mode in reasoning models: the doom loop. A doom loop happens when a model emits a span, then repeats it again and again until the context window runs out. Small reasoning models hit it most on long thinking traces and hard math and coding problems. The usual fixes are stopgaps. Applying repetition_penalty reweights the entire output distribution and can degrade performance. Reinforcement learning can target repetitive looping, but it needs carefully calibrated rewards and costly online rollouts. Our method takes a more targeted approach. We find the exact token that begins a loop, train the model to prefer coherent alternatives at that one position, and leave the rest of the distribution largely intact. The loop almost always starts on a single overtrained token, and in reasoning traces, those are usually interruptives the model has learned to overproduce: "Wait," "So," "Alternatively." Antidoom retrains that first token. It runs on Final Token Preference Optimization (FTPO), a preference-optimization algorithm in the DPO family, built to move a handful of tokens with minimal disturbance to the model otherwise: > Trains only the trailing token of a sequence, mid-generation > Spreads probability across multiple chosen tokens, so it doesn't just replace one overtrained token with another > Uses a KL-like loss in logit space, so training doesn't disturb the rest of the vocabulary The training teaches the model nothing new about math or code. It clears the looping that blocked answers the model could already produce. > Early LFM2.5-2.6B checkpoint: doom-loop rate fell from 10.2% to 1.4% > Qwen3.5-4B: fell from 22.9% to 1% under greedy sampling > Eval scores rose across the board in both cases, attributable entirely to the drop in looping The whole pipeline runs in a few hours. For a 2-4B model, generating the training set takes about two hours on 8 H100s, and training takes one to two hours on a single H100. If your small reasoning models loop on hard prompts, Antidoom recovers the accuracy those loops were costing you. Generation, detection, and the FTPO trainer are all open source.

See how Generative Bionics used our Liquid model in their incredible lower body stack! congrats on this product milestone Daniele Pucci and the team!

What if the same kind of model that generates language could also design biology? Liquid builds foundation models over amino acid tokens, enabling promptable protein design from sequence alone. Check out this discussion between Liquid's CTO Mathias Lechner and Kaeli Kaymak-Loveless, machine learning scientist working on biological applications of LFMs, about designing proteins from sequence alone, engineering an efficient plastic-degrading enzyme, and how the same LFM could point towards new cancer therapeutics.

No more “syntactically correct garbage.” Our CTO, Mathias Lechner, drives home the point that constrained generation, even if it produces valid JSON, can still produce the wrong JSON. A model can return syntactically perfect output and still get the type wrong, invent a field, or pick an enum value that doesn't exist in the schema. It can still fail. This is the problem we built IFStruct for: output that actually satisfies the schema. Read more on the blog → https://lnkd.in/dwHZrcej

Today we release IFStruct, a new benchmark to measure how well models generate structured outputs. Structured output is one of the most common things we ask models to do, and still where they break. Part of the problem is that most benchmarks test with clean, finalized schema. Real requests use plain language, paste an annotated example, switch formats halfway, and slip in constraints like "no code fence" or "no commentary." We built IFStruct to help solve this by presenting requirements in all of those forms: chat requests, bullet lists with explicit paths, raw JSON Schema, annotated JSON or YAML, and ASCII tables. Half are rewritten into natural prose. Scoring is binary. Every field, type, enum, bound, and count right, with no invented keys. The generative design of the dataset means easy production of training data on a held-out set, and the same yes/no check that scores the benchmark can serve as the reward signal to train the model. That's why a 350M model can learn the task well enough to outperform much larger ones. LFM2.5-350M starts at 21.10% and reaches 44.90% after training, ahead of Qwen3.5-4B at 36.25% and granite-4.0-h-tiny at 38.75%. Frontier models near 100%. The test set is available now on Hugging Face. Download the IFStruct → https://lnkd.in/dJvaztHu Read more on the blog → https://lnkd.in/dwHZrcej

A human genome is billions of base pairs. You need an efficient model to handle it. 🧬 Our CTO Mathias Lechner talks with Chief Science Officer Alexander Amini about how Liquid AI is building foundation models for biology.