What people are saying.

Good updates coming to TheStage AI On-device SDK. BTW, new model from Liquid AI LFM2.5-230m runs on my iPhone 16e with 260 tok/s fully on NPU! From our knowledge this is the fastest on-device implementation, moreover with NPU support. Huge respect to Liquid AI team Jeffrey Li Mathias Lechner Yuri Khrustalev Felipe Benavides building so small and smart models! SDK github: https://lnkd.in/eFiqXReG

Last year, I tried journaling for the first time, but it quietly turned into a running log of everything that went wrong, and I felt worse. That bugged me enough to do something about it. Today I'm launching SelfLink on the App Store. 🎉 SelfLink is a journaling app that reads between the lines of your own words and surfaces the good you overlooked, the small wins you didn't notice. And on the days when nothing positive surfaces, a private AI companion helps you find some perspective. Here's the part I care about most: all of it runs 100% on your iPhone. The AI (powered by Liquid AI) runs on your device, not on a server. No account. No cloud. Nothing you write is sent anywhere, and no one reads it, not even me. Your thoughts should stay yours. I built it that way from day one. SelfLink is the newest product from Reality Play. It's free, and there are no ads. If you already journal, or you've always meant to start, I'd love for you to try it :) A link to your better self. 🔗 https://lnkd.in/gRzF2fue

fine-tuned LiquidAI’s LFM2.5-230M on Fable-5 traces and shipped it as GGUF tiny 230M coding-agent model. trained at 4096 ctx. exported Q4_K_M / Q8_0 / F16. runs locally. repo: https://hf.co/AKMESSI/lfm2.5-230m-fable-5

While we eagerly await Fable 5's return, our agentic WebGPU kernel optimization framework kept running. Opus 4.8 picked up where Fable left off, pushing Liquid AI's new LFM2.5 230M to an unbelievable 1,400 tok/s... running locally in your browser. Don't blink or you'll miss it.

You can now train Liquid AI's LFM2-VL in TRL GRPO and RLOO included, with an example script example script: https://lnkd.in/eR5-T9qF thanks https://lnkd.in/eZS49y5x! cc Maxime Labonne

HOLY SPEED BATMAN. THIS THING RIPS ON MY MAC BOOK PRO. Liquid AI drops model and we are racing to test it. #ai #model #24b #moe Right off the rip its destroying my future use of Anthropic and OpenAI. LFG! Mathias Lechner / Ramin Hasani - Kudos. Looking so nice! Link: https://lnkd.in/eyve6hrN

Okay, this is actually insane... You can now run LFM2.5-1.2B-Thinking (a 1.2B parameter LLM from Liquid AI) at over 200 tokens per second directly in your browser on WebGPU! 🤯 Zero install. Fully private. Blazingly fast. Powered by Transformers.js and ONNX Runtime Web

Built on the LFM2 architecture, this 3B parameter model uses transformers and is distributed as safetensors. It's designed to be conversational and efficient, trained for English vision-language tasks. The 'edge' tag hints at its optimized footprint.

@0xTib3rius OSS 20b is the banger in that size range. If you’re looking for a little smaller, but good utility, the LFM2 8b A1b (MOE model) is blazing fast and pretty smart for a little guy (all the LFM models are fantastic) If you need “freedom of expression” look for heretic (or heresy)

KaniTTS2, our text-to-speech model with frame-level position encodings, optimized for real-time conversational AI. ...Full Pretraining Code — train your own TTS model from scratch [https://github.com/nineninesix-ai/kani-tts-2-pretrain](https://github.com/nineninesix-ai/kani-tts-2-pretrain. Highlights: 400M parameter model built on LiquidAI's LFM2 backbone + Nvidia NanoCodec; ~0.2 RTF on an RTX 5080, 3GB VRAM — fast enough for real-time use; Voice cloning with speaker embeddings; Pretrained on \~10k hours of speech data (8x H100s, just 6 hours of training!). Why we're releasing the pretrain code: We want anyone to be able t..

📰 Small LLMs Reveal Surprising Tool-Calling Mastery on CPU — Benchmark Results. A groundbreaking benchmark tests 21 small language models on their ability to judge when to invoke tools, revealing that ultra-compact models like Qwen3:0.6B and LFM2.5:1.2B outperform larger ...#AINews #AI #Teknoloji

Just ran lfm 2.5 from @liquidai locally on my Mac. This was my first experience in running a model locally. It was a great experience with Lm studio and amazing speed and performance by the model. Loved it

🔥 A 1.2B translation model that punches way above its weight. SauerkrautLM-Translator-LFM2.5-1.2B just dropped. This isn't another general-purpose LLM. It's a specialized translator built for one thing — high-quality, nuanced text conversion. What makes

Benchmarking On-Device MLX LLMs with Russet on iPhone 17 Pro and iPad Pro M5. TL;DR: I ran 6 quantized LLMs on Russet which uses Apple's MLX framework on an iPhone 17 Pro and iPad Pro M5, both with 12GB RAM. LFM2.5 1.2B at 4-bit hits 124 tokens/sec on iPad and 70 tokens/sec on iPhone. iPad Pro is 1.2x–2.2x faster depending on model and prompt length, with the gap widening dramatically for longer contexts. More detailed methodology, results (plots included), and discussion in the link.

@kaiapocalypse But it’s incredible that in public benchmarks, LFM2.5 1.2B instruc beats the 8B-A1B MoE in many metrics like MMLU/Pro, GPQA, instruct following and more. 🤌

@tmikov Right now, among models with <1B params, LFM models are 🔥

Wow. Working with small models 1-3B params makes me feel like working with live wires. Tiny changes, big implications. Low key love it. @liquidai Hats off to you folks -- I'm building something on-device and LFM2.5-1.2b punches way above its weight 👊🏾

@LocallyAIApp LFM2 2.6B-Exp-8bit works super fast and efficient in iphone 16 pro