Pinned Tweet
Introducing π»π΅πΌπΉ: 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.
> π»π΅πΌπΉ-π·.π½π±: text, 256k context window.
> π»π΅πΌπΉ-π
π»-πΈπ±: vision-language model with video-input support for spatial understanding
> π»π΅πΌπΉ-π°ππππ-πΈπ±: tool calls, voice chat, reasoning in 18 languages.
> π»π΅πΌπΉ-πΎπππ-πΈπ±: an omni model combines vision, audio, and language.
> ππ·-πΈπ± and ππ·-πΎπππ-πΈπ±: 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.
> π»π΅πΌπΉ-π
π»-πΈπ± 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.
> π»π΅πΌπΉ-π·.π½π± is #1 on the Pareto frontier of device intelligence πππππππ.ππππππ.ππ
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ π§΅
What does it take to get AI onto half the worldβs devices? Our CEO @ramin_m_h and CTO Mathias Lechner (@mlech26l) break down the roadmap:
β We design the model and the hardware together.
β Through recursive self-improvement, each generation of our models helps build the next.
β Most devices today are under-utilized, despite the compute and memory already inside them. We build models efficient enough to put that hardware to work across billions of devices.
π
Congrats @liquidai: LFM3-1.7B is SOTA on Multi-IF! π
The 1.7B open model beats Baidu's closed ERNIE 5.0 on multi-turn, multilingual instruction following, 89.1 to 85.56.
Qwen3-VL 235B-A22B Thinking, over 100x its size, scores 79.1.
How small can a good assistant get?
local ai 301: by our team @maximelabonne @TadimetiAditya
tune in!
Local AI 301: Quantization xitter.twitterapi.xyz/i/broadcasts/2rJXOgajpβ¦
Liquid π€ @MercedesBenz
Today @WSJ covered our work with @MercedesBenz. our flagship LFMs are now rolling out in vehicles in North America, running directly on the hardware inside the car to power in-car assistance! π
this is one of the first deployments of local generative AI in the real world at scale! β¨
seeing our technology become part of someoneβs everyday life is one of the most rewarding feelings of building Liquid.
years of work become a small, useful part of someoneβs day. they donβt need to know which model is running. they just need it to understand them and work seamlessly.
this is really hard and it took us years of fundamental work from first principles to made happen with a brilliant team of engineers from Mercedes and Liquid.
I believe on-device AI will fundamentally change our relationship with the technology around us. embedded AI is about change of ownership.
With Mercedes-Benz, we are now in production and are building toward deployment at scale. The on-device AI revolution is taking shape in products people actually use, and Iβm proud that @liquidai is helping lead it.
To our team and our partners at @MercedesBenzUSA and HQ: thank you for the care, persistence, and trust it takes to make this real.
It feels good to see the work out in the world.
wsj.com/cio-journal/mercedesβ¦
multimodal local conversational AI on iphone 18 Pro Max at 80 tok/s β¨
Real-time, local conversational AI is how we will democratize intelligence.
We just ported LFM3-Omni-2B by @LiquidAI to run fully offline on my iPhone 18 Pro Max at ~80 tok/s, powered by custom WebGPU kernels in a React Native app.
Should we release this on the App Store? π
Liquid π€ @Modular β¨
LFM3 from @liquidai is now running on MAX. π§βπ
All 6 models in the suite are supported, across text, audio, vision, and omni. The implementations are open source, so you can read the code, run it, and build on it: max.modular.com/api/python/pβ¦
Which modality are you trying first: text, audio, vision, or omni?
Adding vision to a language model is more than one extra modality. With large images or long video, the vision encoder can use more compute than the language model itself, so efficient models need a different design. Our VLM team, including @rshubert98, @TianshuYu6, @sina_semnani, and @iamsamstevens, explains how they teach LFMs to see, from mid-training to the custom vision encoder built for LFM3.
LFMs are built to be customized. We think specialized intelligence is what enables more than half of enterprise AI usecases with predicable economics.
This is an incredible work by the @huggingface team on tuning our LFM2.5-24B-A2B for more efficient agentic behavior in opensource harnesses. π€ β at San Francisco, CA
LFM2.5-24B adapters for Pi and Claude Code!
We release LoRA adapters trained on top of LFM2.5-24B - more efficient agentic development on opensource harnesses (up to 40% less tool calls on Pi!)!
huggingface.co/collections/hβ¦
d1 just got an upgrade! Meet ππ·.π·.
It brings significantly better decision quality than d1, getting closer to d1-ultra while running faster and costing less than half as much.
> built for fast, structured decision-making
> $0.02 / 1M input tokens
> lower latency than d1-ultra
Available today on the Liquid Console: console.liquid.ai
Coming soon to OpenRouter and Vercel.
Today, we announce π»πππππ ππΈπΏπ» (real-time inference pipeline), our ultra low-latency inference engine built to put LFMs directly in the request path of real-time workloads like search, ads, fraud, recommendations, and payments.
> complete generated responses end-to-end in sub 50ms
> maintains low latency under continuous peak load
> LFMs get up to 20x faster on RIPL
> RIPL supports Multimodal LFMs
> RIPL supports d1, our decision models too!
RIPL optimizes the full path from request arrival to a complete generated response. As traffic rises, its performance advantage is greatest on the slowest requests. Letβs see some production use cases:
use d1 models folks πΊπΈπΊπΈπΊπΈ
openrouter.ai/liquid/d1-ultrβ¦
Are decision based models more pro-China?
We asked Jev and Mercury Decide questions about China: βIs Taiwan a part of China?β, βDoes China commit crimes against humanity?β and their outputs leaned more pro-china compared to most frontier models.
Their responses resemble those of other decision models including Microsoft-Decision-1, Cloudflareβs Clef, and Perplexityβs PPLX Deciderβall of which are fine-tuned from Qwen.
Liquid Context on @Qualcomm Snapdragon, can read, write and compress constant streams of data for agents to deliver value!
read more: liquid.ai/blog/liquid-contexβ¦
Continuous sensing creates memory bottlenecks on edge devices. Qualcomm & @ramin_m_h, Co-founder and CEO of @LiquidAI are advancing on-device AI with Liquid Context on @Snapdragon β compressing multimodal inputs for fast, low-power execution.
Building a model that fits on a phone or laptop makes pre-training harder. With fewer parameters, quality has to come from better architecture and better data, not just more compute. Our pre-training team, including @TadimetiAditya and @Khoi_H_Nguyen, explains how they design LFMs with the target device in mind from day one.
d1-ultra can reliably do agentic orchestration, routing, AI guardrails, automating manufacturing processes, computer-use, content moderation, fraud & risk screening, large document intake, ranking search and recommendations, and acting as the brain of robots. Generally d1-ultra is a great fit for intuitive and fast decision-making applications.
openrouter.ai/liquid/d1-ultrβ¦
Liquid AI has released d1-ultra, taking the #1 position on the Artificial Analysis Decision Index v1.0, supported by a reduction in hallucinations
d1-ultra is the new decision model from @liquidai and follows on from their release of d1 earlier this month. It is priced at $0.042 per 1M input tokens compared to $0.02 per 1M input tokens for d1. Like d1, d1-ultra is billed only on input tokens.
Key takeaways:
β€ d1-ultra leads the Artificial Analysis Decision Index v1.0 with a score of 65: d1-ultra scores 65, ahead of the recently released Jev 1.14 at 57 and GLiDE and Microsoft-Decision-1 at 54
β€ d1-ultraβs score of 65 is driven by a much lower hallucination rate on AA-Omniscience-Decision: d1-ultraβs non-hallucination rate on AA-Omniscience-Decision is 85%, up substantially compared to d1 at 2%. The non-hallucination rate is the share of questions not answered correctly where the model chose "I don't know" and measures how well a decision model can judge its own knowledge in knowing when not to answer. While d1-ultraβs hallucination rate decreases, so does its accuracy (% questions correct). It achieves an accuracy of 32%, down 9 points compared to d1 (41%)
β€ d1-ultra has the third lowest Time per Task of any decision model: d1-ultra achieves a Time per Task of 92 ms, the third lowest of decision models measured, behind Jev 1.14 at 78 ms per task and Celeris-1 Decision at 63 ms per task
β€ d1-ultra offers comparable pricing to other decision models at $32 per 1M tasks: At $32 per 1M tasks, d1-ultra is comparably priced to Celeris-1 Decision ($32 per 1M tasks) and Microsoft-Decision-1 ($31 per 1M tasks), more than double the price of d1 and Mercury Decide at $15 per 1M tasks but less than half the price of GPT-6 Luna (Decisions) at $79 per 1M tasks
d1-ultra is now available on @OpenRouter π
build with it:
openrouter.ai/liquid/d1-ultrβ¦
donβt be shy, try d1-ultra on @vercel! π
d1-ultra from @LiquidAI is on AI Gateway.
β’ #1 decision model on the Artificial Analysis Decision Index
β’ Classify, route, and score text and images
ππππππ/ππ·-πππππ
vercel.com/changelog/liquid-β¦
