Pinned Tweet
Introducing ๐ป๐ต๐ผ๐น: 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.
> ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ: text, 256k context window.
> ๐ป๐ต๐ผ๐น-๐
๐ป-๐ธ๐ฑ: vision-language model with video-input support for spatial understanding
> ๐ป๐ต๐ผ๐น-๐ฐ๐๐๐๐-๐ธ๐ฑ: tool calls, voice chat, reasoning in 18 languages.
> ๐ป๐ต๐ผ๐น-๐พ๐๐๐-๐ธ๐ฑ: an omni model combines vision, audio, and language.
> ๐๐ท-๐ธ๐ฑ and ๐๐ท-๐พ๐๐๐-๐ธ๐ฑ: 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.
> ๐ป๐ต๐ผ๐น-๐
๐ป-๐ธ๐ฑ 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.
> ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ is #1 on the Pareto frontier of device intelligence ๐๐๐๐๐๐๐.๐๐๐๐๐๐.๐๐
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ ๐งต
Liquid AI retweeted
analyze and sort all of your videos and photos with LFM3!
I'm building the coolest agentic mobile app on @Snapdragon that analyzes and sorts all of your videos and photos on device! This all runs locally on the 8 Elite Gen 5 NPU using @Qualcomm_Dev GenieX and the new @liquidai LFM3 models! Video editing just got a whole lot easier!
Liquid AI retweeted
๐จ ๐ Events in San Francisco alert!
Our researchers at @liquidai keep dropping monster models every other week, so I figured it's only fair that as their SF DevRel, I also host events every week in November ๐ฉโ๐ญ
Here's what we've got coming up!
Liquid AI retweeted
Way to go! Running a complete Omini model (LFM3-Omini-2B) inside a browser is crazy and impressive (and at 280tps on an M4 Max Mac). Imagine what that power could do for Google Chrome, a linux-based Web heavy PC layer.
Liquid AI retweeted
The speed of development becomes your moat.
@ramin_m_h, Co-founder & CEO of Liquid AI, opened FARR Summit with a powerful keynote on building foundation models that run outside the data center, from first principles.
1. Instead of designing the architecture around the teamโs own biases, Liquid AI built an AI system to design it. The result was Liquid Foundation Models.
2. Efficient AI canโt ignore hardware. Hardware and software should be co-designed.
3. The game is product, and the ultimate moat is deep integration.
FARRsummit.com
#FARR #FARRSummit #SFTechWeek
Liquid AI retweeted
I'm building the coolest agentic mobile app on @Snapdragon that analyzes and sorts all of your videos and photos on device! This all runs locally on the 8 Elite Gen 5 NPU using @Qualcomm_Dev GenieX and the new @liquidai LFM3 models! Video editing just got a whole lot easier!
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. ๐ค โ 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/hโฆ
Liquid AI retweeted
โ ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑใฏ
Ryzen7 8700GใฎiGPU๏ผAMD Radeon 780M ๏ผใงๅใใ๏ผ
Q4_K_M ใ llama.cpp๏ผVulkan๏ผใง่ตทๅใงใใใ
่ชญใฟ่พผใฟ pp512
ใป780M๏ผ1240 t/s
ใปCPU๏ผ457 t/s
็ๆ tg128
ใป780M๏ผ47.2 t/s
ใปCPU๏ผ47.7 t/s
ใกใขใชใฏใกใคใณใกใขใชใใๅใใใ262K่ตทๅใง4.9GBใ
โ ๅฎใฏใ24ๆ้ไผใพใชใใ่ชฟๆปไฟใใๆฌฒใใ
ใฏใคใฎXใฏใใใคใๆๆฐๆ
ๅ ฑใซ้
ใใใ
ใฐใฉใStrata Abyssใงๅคงใใใขใใซใๅใใฆใ
ๅ
่ตGPUใง1.7Bใ24ๆ้ๅใใใใใใชๆ
ๅ ฑใๆพใฃใฆ้็ฅใใใใใ
โ ใใ ใ้ ญใฏๆชใ
ๅฐใๅใฃใๆใใ่ชฟๆปใใใฎใพใพไปปใใใใ่ณขใใงใฏใชใใ
Liquid AI่ช่บซใใ็ฅ่ญใ่ฆใไปไบใ่ค้ใชๆจ่ซใซใฏๅใใชใใจๆธใใฆใใใ
ไธๆนใงใใใผใซๅผใณๅบใใจ็นๅฎใฟในใฏๅใใฎ่ฟฝๅ ๅญฆ็ฟใๅๆใซไฝใใใใขใใซใ
โ ใใใใ
LoRAใงใ่ชฟๆปๅฝนใใฎ่ฝๅใ ใไผธใฐใใใใ่ฉฆใใ
1.7Bใงใฉใใพใง่กใใใใฎ็ ็ฉถๅฏพ่ฑกใ
#VRAM้ฃข้ฅๆๆธๆ
Introducing ๐ป๐ต๐ผ๐น: 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.
> ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ: text, 256k context window.
> ๐ป๐ต๐ผ๐น-๐
๐ป-๐ธ๐ฑ: vision-language model with video-input support for spatial understanding
> ๐ป๐ต๐ผ๐น-๐ฐ๐๐๐๐-๐ธ๐ฑ: tool calls, voice chat, reasoning in 18 languages.
> ๐ป๐ต๐ผ๐น-๐พ๐๐๐-๐ธ๐ฑ: an omni model combines vision, audio, and language.
> ๐๐ท-๐ธ๐ฑ and ๐๐ท-๐พ๐๐๐-๐ธ๐ฑ: 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.
> ๐ป๐ต๐ผ๐น-๐
๐ป-๐ธ๐ฑ 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.
> ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ is #1 on the Pareto frontier of device intelligence ๐๐๐๐๐๐๐.๐๐๐๐๐๐.๐๐
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ ๐งต
Liquid AI retweeted
Same model. Same accuracy. 39% fewer tool calls.
We RL-trained a LoRA so LFM2.5-24B-A2B works better inside the pi harness: it solves as many tasks, with far fewer steps.
@LiquidAI_ base + @badlogicgames's pi + TRL AsyncGRPO + ~24h on 1 GPU
huggingface.co/collections/hโฆ
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/hโฆ
Liquid AI retweeted
I saw @helloiamleonieโs recipe for fine-tuning a @LiquidAI decision model on text and wanted to try it with images too.
Started with d1-3B but also added d1-2B as well, which was just released.
Given a website screenshot, a task, and candidate page elements, the decision model selects which element to interact with next.
After one epoch on Mind2Web, using one RTX PRO 6000 per model on @modal:
โข d1-3B: 44% โ 91% accuracy, ~44 min training
โข d1-2B: 51% โ 96% accuracy, ~20 min training
Iโve shared the training, evaluation, and endpoint-serving code, plus a video walkthrough:
drchrislevy.com/blog/liquid-โฆ
Liquid AI retweeted
We've just made a breakthrough: the first true omni model running 100% locally in a browser... fully end-to-end, with no separate ASR or TTS models needed.
With custom WebGPU kernels, we're able to run LFM3-Omni-2B at ~280 tok/s on my M4 Max, accepting visual, audio, or text and producing interleaved text and audio tokens. Just 90 ms to the first token and 140 ms to the first audio.
The future of on-device multimodal AI agents is going to be wild.
Liquid AI retweeted
@liquidai team is on ๐ฅ, but the big thing is still under cooking. Stay tuned ๐ฆ๐บ
Liquid AI retweeted
What if a child could watch/record an educational video and ask questions about it without having to send any data to the cloud.
This is the use case Iโve been working on for sometime and thanks to @liquidai new LFM3-Omni-2B, Iโve been able to make a decent prototype with a single model. Itโs a compact yet a very capable multimodal model.
โขย Kids could pause and ask questions about the video by text or voice,
โขย Highlight a part of the video to get an explanation
Everything runs on device.
On-device models have a lot of potential in education, especially where connectivity is low or privacy is a concern.
๐ป๐ต๐ผ๐น-๐พ๐๐๐-๐ธ๐ฑ is our first omni model, built for cross-modal interactions that require understanding visual and auditory information together. It combines vision, audio and language in a single model that responds in text, speech or both.
Ask a spoken question about an image, or provide a video with its soundtrack and ask something that requires understanding both. The model processes visual and auditory information jointly.
It also combines visual context with spoken instructions to generate structured tool calls, helping assistants act on what users show and tell them.
> In our evaluations, it scores 70.1 on Daily-Omni, which tests video and audio understanding, compared with 61.3 for Qwen2.5-Omni-7B.
At 2.2B parameters, it brings multimodal agents on-device that use visual and auditory context to guide users, respond to voice commands and act through tools. Developers can specialize this foundation model for personalized education, everyday assistance, in-car experiences, digital interfaces, robotics, and more.
> ๐ป๐ต๐ผ๐น-๐พ๐๐๐-๐ธ๐ฑ model card: huggingface.co/LiquidAI/LFM3โฆ
7/
Liquid AI retweeted
a year ago, an interviewer called me naive.
all because i said customers would fine-tune open models instead of using frontier APIs. (true story)
this week, we released LFM3 at @liquidai:
small open models built to be specialized.
as a demo, I fine-tuned our 2B vision model to spot defects in drone photos of wind turbines.
before:
LFM3-VL-2B and GPT-6 Luna named the defect correctly in under 10% of images.
after:
a few hours of fine-tuning on one GPU made it jump to 67%.
frontier models are generalists.
your domain-specific problem needs a specialist.
how you can do it too:
1. pick your domain-specific, narrow task and evaluate if a frontier model is sufficient. If not:
2. choose the LFM3 model that fits your data.
3. curate your own examples from your own proprietary data
4. fine-tune in a few hours, even on a single GPU.
5. rerun the evals and deploy anywhere without having to call a cloud API.
Fine-tuning notebooks:
github.com/Liquid4All/cookboโฆ
Wind turbine project: github.com/Liquid4All/cookboโฆ
what task would you specialize a model for first?
Liquid AI retweeted
Meet LFM3-Omni-2B
๐ค Made with AI
Introducing ๐ป๐ต๐ผ๐น: 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.
> ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ: text, 256k context window.
> ๐ป๐ต๐ผ๐น-๐
๐ป-๐ธ๐ฑ: vision-language model with video-input support for spatial understanding
> ๐ป๐ต๐ผ๐น-๐ฐ๐๐๐๐-๐ธ๐ฑ: tool calls, voice chat, reasoning in 18 languages.
> ๐ป๐ต๐ผ๐น-๐พ๐๐๐-๐ธ๐ฑ: an omni model combines vision, audio, and language.
> ๐๐ท-๐ธ๐ฑ and ๐๐ท-๐พ๐๐๐-๐ธ๐ฑ: 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.
> ๐ป๐ต๐ผ๐น-๐
๐ป-๐ธ๐ฑ 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.
> ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ is #1 on the Pareto frontier of device intelligence ๐๐๐๐๐๐๐.๐๐๐๐๐๐.๐๐
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ ๐งต
Liquid AI retweeted
ran d1-2B on 3,080 real bank tickets on a cpu. when it said 90% sure it was right 98.7% of the time. full results in my thread ๐
nitter.gq/humayun_x/status/21158โฆ
Liquid AI retweeted
LFM3-VL-2B Q4
5090 ็ฐๅขใงๅธธๆ 470tok/s ใใใๅบใฆใฆใฑใ
ๆฐใฉใชใผใงๆธใใใใช่ปฝใฟในใฏใชใๅ
จ็ถๅ้กใชใใใ
ไธๅฟใใใชในใไฝใฃใฆใฟใฆ่ฒฐใฃใใใฉใจใฉใผใงใพใจใใซ่ตทๅใงใใชใใไฟฎๆญฃใใใฆใฟใฆใใ ใใ ใฃใ
ใพใใใใใใซใญ
BF16 ใ่ฉฆใใฆใฟใใฃใใช
ใใฏ LFM3 omni ใใข
ๅ็ปใญใฃใใทใงใณ้ใ๐คฏ
๏ผ็งใฎๅ็ปใ๏ผ็งใงใญใฃใใทใงใณใงใใ
ใพใ ๅฐใใฎๅ็ปใงใใ่ฉฆใใฆใชใใใฉใใชใขใซใฟใคใ ็ณปใฎ็จ้ใซใไฝฟใใๅฏ่ฝๆงใใใช๏ผ๐ค
้ณๅฃฐ็ๆใฏโฆใใ๐
ๆฅๆฌ่ชใฎใใใใจๆใฃใใใฉ่ฑ่ชใงใๅชๆ่ฉ ๅฑใใฆใ
ใพใ 2B ใ ใใญ
TTS ใฏๅ
ใ
่ปฝ้ใขใใซ่ฒใ
ใใใใ็นใซๅฐใใชใใญ
ใใจใๅ็ปใญใฃใใทใงใณใฏใใฌใผใ ็ปๅใ ใ่ฆใฆใใฎใใใใจๆใฃใใใฉใใกใใใจ้ณๅฃฐใ่ใใฆใใ
ๆๅพใฎใฏๅ
ฅๅๅ็ป
Liquid AI retweeted
pointed lfm3-vl-2b at the whole san fernando valley overnight. 16,692 aerial tiles, one rtx 3060, 5.5 hours.
it found 62,873 pools and drew a box around each one.
not a crazy idea, just โfind the poolsโ. it misses about 1 in 5 and boxes a roof now and then. but itโs a 2B model with zero finetune.
imagine what you can do with this. palantir at home lmao
2/ LFM3-VL-2B
64 slots x 16K. ~143 tok/s on one request, ~583 tok/s total with all 64 describing images at once. 11.5GB VRAM.
config:
llama-server
-m LFM3-VL-2B-Q8_0.gguf
--mmproj mmproj-LFM3-VL-2B-F16.gguf
-ngl 99 -c 1048576 -fa on --jinja
-np 64
--cache-type-k q8_0 --cache-type-v q8_0
-b 2048 -ub 1024
--temp 0.1 --top-p 1.0 --top-k 50 --min-p 0.0 --repeat-penalty 1.1
--reasoning on --no-reasoning-preserve
Liquid AI retweeted
Congrats to @liquidai on LFM3! Now in Nativ ๐
Text, vision, audio and omni, all around 2B ยท fully on-device.
Ask your Mac something out loud with the Wi-Fi off, and it usually answers out loud in under a second.
On an M5 Max at 4-bit:
๐๏ธ an hour of audio turned into text in about a minute
๐ Omni answers a question about a photo in 0.15 s
๐ฌ and about a short clip in half a second
๐พ under 3.3 GB at peak
Liquid's first omni model, and nothing you say leaves your Mac.
Download for macOS ๐ป
runnativ.ai
Collections ๐ช
huggingface.co/collections/Lโฆ
Github repo ๐
github.com/Blaizzy/nativ
Liquid AI retweeted
Mid-read, you ask out loud: "Compare L4-L5 with L3-L4, and show L4 from the side."
SpineLab Voice measures both discs and turns the 3D spine to match. Fix a landmark, ask again, and it re-measures from your fix.
Liquid AI's LFM3-Audio-2B hears the request and calls the tools.
Liquid AI retweeted
LFM3-1.7B just read the whole of Pride and Prejudice on a M5 Pro using MLX.
It read 100k tokens in about 28 s (sped up 8ร in the clip), then answered 100 questions back to back in 27 s, including finding a password hidden somewhere in the book.
huggingface.co/alexwengg/LFMโฆ
Introducing ๐ป๐ต๐ผ๐น: 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.
> ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ: text, 256k context window.
> ๐ป๐ต๐ผ๐น-๐
๐ป-๐ธ๐ฑ: vision-language model with video-input support for spatial understanding
> ๐ป๐ต๐ผ๐น-๐ฐ๐๐๐๐-๐ธ๐ฑ: tool calls, voice chat, reasoning in 18 languages.
> ๐ป๐ต๐ผ๐น-๐พ๐๐๐-๐ธ๐ฑ: an omni model combines vision, audio, and language.
> ๐๐ท-๐ธ๐ฑ and ๐๐ท-๐พ๐๐๐-๐ธ๐ฑ: 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.
> ๐ป๐ต๐ผ๐น-๐
๐ป-๐ธ๐ฑ 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.
> ๐ป๐ต๐ผ๐น-๐ท.๐ฝ๐ฑ is #1 on the Pareto frontier of device intelligence ๐๐๐๐๐๐๐.๐๐๐๐๐๐.๐๐
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ ๐งต