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/ π§΅
Nathan Ranchin retweeted
Want to try weight-space programming? Today we're releasing dspy-wsp, a prototype weight-space optimizer that works like any other @DSPyOSS optimizer.
Using dspy-wsp, and about ~$5-10 of Astra, we improved every decision model we tested. cmpnd.ai/blog/weight-space-pβ¦
βblazingly fast audio codecβ -> β‘οΈππ€
Introducing π»ππππππ²ππππ, our ~74M-parameter streaming audio codec powering LFM3-Audio-2B and LFM3-Omni-2B.
An audio codec turns speech into tokens a language model can predict, then converts those tokens back into audio. We built ours around two requirements: tokens that are easy to predict and fast conversion back into speech.
On a single CPU thread on an Apple M4 Max, LiquidCodec runs at approximately:
> 361Γ real time for encoding
> 87Γ real time for decoding
Both measurements use FP32 precision.
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.
Nathan Ranchin retweeted
on Thursday I will be speaking at @sgl_project summit about inference on heterogenous chips; come say hi π€
local ai 301: by our team @maximelabonne @TadimetiAditya
tune in!
Local AI 301: Quantization teapawt.org/i/broadcasts/2rJXOgajpβ¦
Nathan Ranchin retweeted
LFM3 Omni vs GPT6 Astra - chess edition?
we put the LFM (@liquidai) model that @manoletre finetuned for chess, head-to-head up against GPT6-Astra.
Astra ran with reasoning enabled and took on average over 100x longer than LFM on every move.
@liquidai FTW π
I adapted @liquidai's d1-omni-600b to play chess and listen to my voice and I haven't been able to beat it.
It was the final result of a lot of experimentation and a lot of fun.
To me the highlight was after training, asking claude to try experiments to reduce the time the model needed to make a move, peak memory usage, size of the model file, etc *while* keeping chess performance constant.
Claude found some cool stuff. For example that only ~2k tokens were being used and that we could get rid of a large portion of the embeddings matrix.
Play against the model on my website! ππ½ It runs completely locally on your browser's tab.
Nathan Ranchin retweeted
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¦
Nathan Ranchin retweeted
Please just ask Astra to use Antidoom instead, it's 100x cheaper and faster.
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! π§΅
Blog: rikiyat.github.io/blog/weighβ¦
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-spβ¦
github.com/RikiyaT/lm_undβ¦
huggingface.co/Qwen/Qwen3.5-0β¦
huggingface.co/RikiyaT/Qwen3.β¦
Two billion parameters, three ways to act: planning, navigation, and manipulation - excited to share some of our latest work along with our new model family. πππ
Replying to @liquidai
π»π΅πΌπΉ-π
π»-πΈπ± can also serve as a foundation for robotics. We explore three adaptations spanning the robotics hierarchy, from tool-based task planning to navigation and manipulation.
Tool orchestration: ground objects by pointing, measure them with depth tools, call motion planners, and check outcomes in new camera images. Pair it with a fine-tuned d1-2B skill monitor to detect failures and guide recovery.
Navigation: translate route instructions and a sequence of image observations into visually grounded waypoint trajectories. The model achieves 53% success on R2R and 43% on RxR in unseen scenes under strict continuous-control evaluation with wall sliding disabled.
Manipulation: turn images, language, and robot state into multi-step action chunks in one forward pass. Jointly trained across 15 embodiments, the resulting VLA achieves 96.2% average success across the four LIBERO suites and 45.7% on RoboCasa-GR1 tabletop tasks.
The manipulation VLA generates action chunks in under 50 ms on an @nvidia Jetson AGX Orin using TensorRT in FP16, enabling real-time action generation on embedded hardware.
10/
Nathan Ranchin retweeted
SGLang has Day-0 support for LFM3 from @liquidai! LFM3 is a family of small models built to be specialized and run anywhere.
Β· New tri-hybrid architecture: LFM3-1.7B beats LFM2.5-2.6B with 35% fewer parameters
Β· One backbone for every modality: text, vision (now with video), audio, omni, and decision
Β· Built for the edge: ~1 GB at 4-bit, 256K context for on-device agents
Β· Fine-tunes on a single GPU
Get started with SGLang π
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§΅
Nathan Ranchin retweeted
LFM3: new generation with six edge models
I'm super happy with LFM3-1.7B, it's an absolute (agentic) monster
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§΅
Been cooking at @liquidai π§βπ³
LFM3-1.7B: best average across 29 text benchmarks, against models up to 5.7x its size.
And thatβs just text. Vision, audio, omni and decision models all ship today on the same backbone. Proud of the team π₯
Try it yourself: huggingface.co/collections/Lβ¦
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§΅
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.
β‘οΈ
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:
Nathan Ranchin retweeted
Perplexityβs Decider v1.2 led the 9 shared models on our test too, scoring 67.1%.
Liquidβs d1-ultra is not on ClaimBench. On the same questions, it scored 77.1%, 10.0 points ahead of Decider v1.2.
zekiworks.github.io/analysisβ¦
pplx-decider-v1.2 is the highest scoring decision model on ClaimBench.
Try pplx-decider-v1.2 and pplx-decider-flash-v1.2 on OpenRouter or Perplexity API Console.
pplx.ai/openrouter-v1.2
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.
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.
You should try @liquidai's d1 models as well.
d1-3B vs Vega 4B on M4 Max: 3.4Γ faster. More accurate. Fewer parameters.
Available here: huggingface.co/LiquidAI/d1-3β¦ and huggingface.co/LiquidAI/d1-oβ¦






























