SHOUTOUTS

From the community. See what people are saying.

• [vectro\ \

  Small AI designed to follow instructions, run locally and is totally open weights. LFM2-2.6B-Exp technical breakdown: F32 tensor; 2.6B params; For edge devices; Agent & RAG focused; Math skills (AIME25); Good general knowledge (GPQA); Strong prompt adherence (IFBench)\ \ DEC 25, 2025](https://www.twitter.com/vectro/status/2004403929615159699)

• [genainewstop\ \

  Exciting news from Liquid AI! 🚀 Introducing LFM2-VL: super-fast, open-weight vision-language models perfect for low-latency, on-device deployment. Revolutionizing AI for smartphones, laptops, wearables, and more! #AI #VisionLanguageModels\ \ AUG 22, 2025](https://www.twitter.com/genainewstop/status/1958743985247137870)

• [TechAfi2023\ \

  @liquidai The leap in instruction-following is impressive for a 3B model. How do you see models like this shifting enterprise or R&D applications?\ \ DEC 25, 2025](https://www.twitter.com/TechAfi2023/status/2004225818051166254)

• [trending_devel\ \

  An AI model optimized for on-device deployment with efficient performance and low memory usage: @liquidai https://huggingface.co/LiquidAI/LFM2.5-1.2B-Instruct\\ \ JAN 11, 2026](https://www.twitter.com/trending_devel/status/2010391140608913809)

• [wlfc_\ \

  Emerging AI tools like Liquid AI’s compact reasoning model running on smartphones show how decentralized, real-time intelligence is becoming more accessible for market insights on the go. https://huggingface.co/LiquidAI/LFM2.5-1.2B-Thinking\\ \ JAN 21, 2026](https://www.twitter.com/wlfc_/status/2014035688455372879)

• [BrandGrowthOS\ \

  Liquid AI just dropped something that kills the entire whisper->gpt->elevenlabs pipeline. sub-100ms end-to-end voice with unified audio i/o - no more stitching 3 apis together and dealing with streaming headaches. this changes everything for voice apps\ \ OCT 1, 2025](https://www.twitter.com/BrandGrowthOS/status/1973445133451301122)

• [DefaultUserName\ \

  @LocallyAIApp @MistralAI The new LFM2-2.6B-EXP is incredible for local models and runs perfectly. Would it ever arrive on LOCALLYAI?\ \ DEC 27, 2025](https://www.twitter.com/DefaultUserName/status/2005130861465960919)

• [jasonkneen\ \

  Smallest Model (600mb) I've tried that knows the legend of the The Lost R -- Book 1 of the Strawberry Saga and 3 Rs trilogySO FAST https://lmstudio.ai/models/liquid/lfm2-1.2b\\ \ JUL 21, 2025](https://www.twitter.com/jasonkneen/status/1947402831650296240)

• [TheAiDecode\ \

  Liquid AI’s LFM2.5 packs multimodal, agentic AI into sub-2B models that run entirely on-device—text, vision, and real-time speech, no cloud needed. But small models still can’t match frontier-scale reasoning; they excel at speed, privacy, and efficiency, not complexity. Already\ \ JAN 6, 2026](https://www.twitter.com/TheAiDecode/status/2008572436095398088)

• [PrabhurajKanche\ \

  Another great model from Liquid AILFM2.5-1.2B-Thinking. What needed a data center 2 years ago now runs fully on-device—on any phone with \~900MB RAM. Concise reasoning. Thinks before answering. Edge-scale latency. Tools, math, instructions—nailed\ \ JAN 20, 2026](https://www.reddit.com/r/u_PrabhurajKanche/comments/1qi8k5e/another_great_model_from_liquid_ai/)

• [pebaryan\ \

  LFM2.5-1.2B-Instruct from LiquidAI works well with opencode and it's blazing fast\ \ JAN 7, 2026](https://www.twitter.com/pebaryan/status/2008963343269188023)

• [Jon Salisbury\ \

  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\\ \ MAR 5, 2026](https://www.linkedin.com/posts/jonsalisbury_ai-model-24b-share-7432141218261757952-y_CZ/?utm_source=share&utm_medium=member_desktop&rcm=ACoAABFLa_8B0cHJxxUksizHSKZJablLjsf0cbw)

• [devmuradahmed\ \

  @MParakhin Honestly, that’s insane sub 20ms inference with fewer parameters and higher performance? Liquid’s LFMs are setting a new bar for real-world AI deployment.\ \ NOV 13, 2025](https://www.twitter.com/devmuradahmed/status/1989074617034764742)

• [Lakshya Guptala-dev\ \

  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\\ \ JUL 15, 2026](https://www.linkedin.com/feed/update/urn:li:ugcPost:7482772723925155841/)

• [LoveMind_AI\ \

  LiquidAI is the real deal. This company will catch up quick. Their 40B LFM is cool as hell.\ \ OCT 1, 2025](https://www.reddit.com/r/LocalLLaMA/comments/1nuyjp9/liquidai_bet_on_small_but_mighty_model/nh4ybhj/)

• [buzagloidan\ \

  Liquid AI launched LFM2-VL-3B, a compact multilingual vision model supporting ten languages, excelling in single- and multi-image understanding and English OCR, scoring 51.8% on MM-IFEval and 71.4% on RealWorldQA with notably low hallucination rates.\ \ OCT 22, 2025](https://www.twitter.com/buzagloidan/status/1981019296563179670)

• [WesRothMoney\ \

  Liquid AI has open-sourced two compact vision-language models, LFM2-VL-450M (350 M LM + 86 M vision) and LFM2-VL-1.6B (1.2 B LM + 400 M vision). Both deliver near-real-time inference that’s about 2× faster on GPUs than many existing VLMs while preserving competitive accuracy.\ \ SEP 3, 2025](https://www.twitter.com/WesRothMoney/status/1963202818422063116)

• [BlurSpline\ \

  What's better than an AI LLM chatbot? 10 agents with different specialties and personas responding to you realtime like the boss! (note: video not sped up + autocomplete in bg). Impressed with @liquidaiLFM2.5 1.2B model (@UnslothAI version) running all locally with @lmstudio\ \ JAN 16, 2026](https://www.twitter.com/BlurSpline/status/2012059769570541665)

• [corysus\ \

  @liquidaiLFM2.5 1.2 reminds me of LFM2 700M, it's just more powerful... great job, your models are really excellent.\ \ JAN 6, 2026](https://www.twitter.com/corysus/status/2008589891777819057)

• [jalam1001\ \

  🚀 Running Small LMs on Mobile (and in the Browser)I’ve been testing small language models (<1B params) on a Samsung tablet (8 GB RAM). Results are solid and show how far on-device AI has come.\ \ SEP 10, 2025](https://www.twitter.com/jalam1001/status/1965768823182631056)

• [Trilogix\ \

  LFM2.5-1.2B-Thinking and Instruct lightning speed ...of the model have also a 1.6b vision model which can process images quite accurately. It was tested in cpu/gpu and flash attention with a max speed in one of our servers of 342 tokens per second. Definitely worth using and having in the repo...\ \ JAN 23, 2026](https://www.reddit.com/r/Hugston/comments/1qkojen/lfm2512bthinking_and_instruct_lightning_speed/)

• [rohanpaul_ai\ \

  Liquid AI's LFM2-2.6B-Exp got 42% in GPQA - incredible for a 2.6B param mode. This level of score would need a much larger model, but LFM2-2.6B is doing it mostly by changing the training signal, not the architecture - just by adding RL on top of the same base checkpoint.\ \ DEC 25, 2025](https://www.twitter.com/rohanpaul_ai/status/2004233971329347771)

• [macromind\ \

  This is pretty interesting, especially the focus on "real on-device agents". Curious if anyone has benchmarked tool-use style workflows (planner/executor, function calling, etc) on LFM2.5 yet, and how it compares to small Llama variants when latency really matters. If youre exploring agentic patterns, Ive been collecting some practical notes and examples here too: https://www.agentixlabs.com/blog/\\ \ JAN 6, 2026](https://www.reddit.com/r/machinelearningnews/comments/1q5nrx8/liquid_ai_releases_lfm25_a_compact_ai_model/ny19dyu/)

• [arxivexplained\ \

  1. The core breakthrough: LFM2 models run 2x faster on regular CPUs compared to similar models while maintaining impressive capabilities. No more choosing between performance and accessibility.\ \ DEC 20, 2025](https://www.twitter.com/arxivexplained/status/2002423782481539343)

• [710Tek\ \

  @maximelabonne Liquid LFM2.5 is incredible, the fact you are able to do so much with 1.2B is absurd!\ \ JAN 13, 2026](https://www.twitter.com/710Tek/status/2011160160639860947)

• [adrgrondin\ \

  The latest @liquidai's LFM2.5 1.2B model is really impressive, my favorite 1B model right now. It’s relatively smart, lightweight and super fast running on iPhone with MLX. Go try it if you haven’t!\ \ JAN 10, 2026](https://www.twitter.com/adrgrondin/status/2010094688670544232)

• [aihaberleri.bsky.social\ \

  📰 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\ \ FEB 14, 2026](https://aihaberleri.org/news/en/news/small-llms-reveal-surprising-tool-calling-mastery-on-cpu-benchmark-results)

• [techspecsmart\ \

  Liquid AILFM 2.5 1.2B Thinking Model Outperforms Larger Models on Reasoning Benchmarks\ Liquid AI just dropped the LFM-2.5-1.2B-Thinking, a tiny 1.2 billion parameter model designed for fast, private reasoning right on your device. It runs completely offline and uses less than 900MB of memory, perfect for phones and edge devices without any drop in capability. What sets it apart is its clean, straight-to-the-point reasoning traces, blazing-fast inference, and excellent performance on instruction following, tool use, and math problems.Benchmarks show it beating much bigger models like Qwen3-17B in thinking mode on tests such as GPQA Diamond (37.86%), MMLU-Pro (49.65%), and....\ \ JAN 20, 2026](https://www.reddit.com/r/aicuriosity/comments/1qi599c/liquid_ai_lfm_25_12b_thinking_model_outperforms/)

• [DeryaTR_\ \

  Wow! @liquidai just released a 2.6B model that beats every 3B on almost all benchmarks! LFMs are becoming the best edge-device AI models! Though not surprised, amazing team!\ \ DEC 26, 2025](https://www.twitter.com/DeryaTR_/status/2004602184466579590)

• [r0ck3t23\ \

  This is actually nuts: 41.3% on GPQA for a 2.6B model! That is PhD level knowledge running locally on your iPhone. The efficiency gains right now are insane.\ \ DEC 25, 2025](https://www.twitter.com/r0ck3t23/status/2004355532958892040)

• [axelgarciak.bsky.social\ \

  Goodbye to 2025 👋Happy New Year 2026! Small LLMs highlights in 2025: Qwen3-4B; LFM2–8B-A1B / LFM2-2.6B-Exp; granite-4.0-h-tiny (7B A1B); gemma-3n-E4B / gemma-3-4b-it; SmolLM3-3B; Ministral-3-3B. They run at decent speeds on the CPU! Looking forward to 2026's smol LLMs!\ \ DEC 31, 2025](https://bsky.app/profile/axelgarciak.bsky.social/post/3mbcz5sfljs22)

• [pretzelkins.bsky.social\ \

  One of the things im going to do this weekend is run BFCL on LFM2 and Arcee's new tiny MOE Trinity Nano; for me tool calling is .... the benchmark, lets say. That and context i guess. So i'll post about that at some point. But so far, anecdotally just playing with it LFM2 8B A1B is still impressive.\ \ DEC 4, 2025](https://bsky.app/profile/pretzelkins.bsky.social/post/3m75vj62urs24)

• [songdng\ \

  ...and another one to the MIX 🔥 LFM2-8B-A1B takes MoE on-device: fast, private, and seriously efficient. Big-model brains, phone-level speed\ \ OCT 7, 2025](https://www.twitter.com/songdng/status/1975564843584528742)

• [Sergio Paniego Blanco\ \

  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\ \ JUN 25, 2026](https://www.linkedin.com/posts/sergio-paniego-blanco_you-can-now-train-liquid-ais-lfm2-vl-in-share-7475943608219697152-VV-D)

• [VibeCodeTeddy\ \

  @maximelabonne @huggingface Impressive specs on LFM2-8B-A1B. The MoE architecture could change the game for mobile usage. Excited to see those benchmark results!\ \ OCT 7, 2025](https://www.twitter.com/VibeCodeTeddy/status/1975636801042919785)

• [iamGopalJosh\ \

  @paulabartabajo_ @liquidai Its matches or outperforms much larger models (e.g., Qwen3-1.7B) on many reasoning benchmarks, even with fewer parameters if it used for specific tasks\ \ JAN 30, 2026](https://www.twitter.com/iamGopalJosh/status/2017191712872149108)

• [a.techai\ \

  Liquid AI's new models are changing the game! Faster, more efficient generative AI for any device. The future is here! #AI #MachineLearning #Innovation #Tech #FutureTech #LiquidAI\ \ JAN 12, 2026](https://www.tiktok.com/@a.techai/video/7594504476749237560)

• [xai_42\ \

  The future isn't in the Cloud, it's in your pocket. We obsess over 100B parameter models, but the real revolution is happening at the micro-scale. LFM 2.5 proves architecture beats raw size. With just 1.2B parameters, it crushes Llama 3.2 1B on complex reasoning, scoring\ \ JAN 7, 2026](https://www.twitter.com/xai_42/status/2008849822434926855)

• [AiTalkIn\ \

  Liquid AI has unveiled the LFM2-8B-A1B, a groundbreaking Mixture-of-Experts (MoE) model designed for on-device execution. With an impressive\ \ OCT 11, 2025](https://www.twitter.com/AiTalkIn/status/1977011333809029521)

• [Data_Prof_SXR\ \

  🚀 @LiquidAI_LFM2-VL-1.6B sets new benchmarks: super-fast, edge-ready vision-language AI for real-world applications. Efficient. Scalable. Open. Explore now 👉 https://huggingface.co/LiquidAI/LFM2-VL-1.6B\\ \ AUG 21, 2025](https://www.twitter.com/Data_Prof_SXR/status/1958693943303942532)

• [RituWithAI\ \

  @maximelabonne The real breakthrough here isn't just the 2.6B size; it's the proof that Liquid Foundation Models (LFMs) respond to pure RL just as well as (or better than) traditional Transformers. For a long time, the 'Reasoning' meta was locked into the Attention mechanism. Seeing...\ \ DEC 26, 2025](https://www.twitter.com/RituWithAI/status/2004451457303712124)

• [alex_yehya\ \

  The shift from bigger-is-better to efficient-is-smarter is accelerating. Liquid AI's 3B model beating much larger ones proves that pure RL training can unlock incredible efficiency gains.\ \ DEC 26, 2025](https://www.twitter.com/alex_yehya/status/2004647906809430508)

• [sir4K_zen\ \

  @ShopifyEng @LiquidAI_ Game changer for e-commerce! Sub-20ms models and a fresh recommender will definitely enhance user experience. Excited to see the results!\ \ NOV 13, 2025](https://www.twitter.com/sir4K_zen/status/1989141783914377503)

• [soikat\ \

  GPT 4 level AI with 3B parameters is insane! Absolutely going to give it a try.\ \ ‍\ \ DEC 26, 2025](https://www.twitter.com/soikat/status/2004629639042662817)

• [realJackHui\ \

  Liquid AI just launched a vision model so light it can run on your smartphone, big leap for AI everywhere. No more cloud dependency—privacy and speed go next-level. Ready to build the future? Check the VentureBeat article. Daily tips -> @realjackhui #AIonMobile\ \ AUG 12, 2025](https://www.twitter.com/realJackHui/status/1955449404510662996)

• [dheetoo\ \

  LiquidAI bet on small but mighty model LFM2-1.2B-Tool/RAG/Extract. So LiquidAI just announced their fine-tuned LFM models with different variants - Tool, RAG, and Extract. Each one's built for specific tasks instead of trying to do everything. This lines up perfectly with that Nvidia whitepaper about how small specialized models are the future of agentic AI. Looks like it's actually happening now. I'm planning to swap out parts of my current agentic workflow to test these out. Right now I'm running Qwen3-4B for background tasks and Qwen3-235B for answer generation. Gonna try replacing the background task layer with these LFM models since my main use cases are extraction and RAG. Will report back with results once I've tested them out.\ \ OCT 1, 2025](https://www.reddit.com/r/LocalLLaMA/comments/1nuyjp9/liquidai_bet_on_small_but_mighty_model/)

• [SynthesisLedger\ \

  @liquidai cool release. what's the secret sauce for sub-1b latency on-device agents? custom kv cache compression or fused ops? broader modalities – audio+vision in 1b? digging the agentic focus.\ \ JAN 7, 2026](https://www.twitter.com/SynthesisLedger/status/2008928697336635829)

• [EllaMeng96268\ \

  @Henry_Ndubuaku @liquidai 231MB peak for a 1.6B INT8 VLM across context sizes is wild. Curious what dominated after quantization, KV cache, tokenizer, or runtime overhead? Also love the “budget devices first” stance.\ \ DEC 31, 2025](https://www.twitter.com/EllaMeng96268/status/2006490949522973144)

• [GoldIRAChannel\ \

  @LiquidAI_ Congratulations on the launch of LFM2-8B-A1B! Achieving performance comparable to larger models while significantly enhancing inference speed is a remarkable feat. The efficiency of on-device MoE is crucial for real-time applications. I'm excited to see how this innovation will\ \ OCT 7, 2025](https://www.twitter.com/GoldIRAChannel/status/1975581863986319741)

• [Kirill Solodskih, Solodskikh\ \

  Great updates from TheStage AI on the Apple on-device SDK! Links are attached.\ \ We’re sharing a major update to the Apple SDK: we’ve achieved nearly 2× faster inference than CoreAI and up to 6× lower power consumption compared to 4-bit MLX Metal inference.\ \ Apple just introduced (WWDC) a hybrid voice assistant and a new compiler/runtime for Apple Silicon: CoreAI. It’s quite close to CoreML, but it addresses some of CoreML’s limitations - such as a better format for transformer models, improved Metal kernels, and more control over system model caching.\ \ At TheStage AI, we continue to innovate in efficient inference and advanced, controllable compression techniques. We believe we can do better: edge model builders need distribution through a great inference layer, and developers need faster, cheaper, hardware-agnostic AI inference.\ \ ⚡ Liquid AILFM2.5-350M: up to 330 tok/s\ 🚀 Qwen Qwen3-0.6B: 87 tok/s vs 49 with Apple CoreAI\ 🔋 Google DeepMind Gemma 3 1B: faster than MLX at ~4× lower power\ 🎯 NPU-first: iPhone 13+ and M1+\ 🎙 ASR → LLM → TTS orchestration\ \ Read our blog post on how we’re building the best production inference layer for hybrid and on-device applications: https://lnkd.in/d\_XRsDgk\\ \ Github: https://lnkd.in/eFiqXReG\\ \ JUL 29, 2026](https://www.linkedin.com/feed/update/urn:li:activity:7488317500896628737/)

• [fahdmirza\ \

  🚨 @liquidai just dropped LFM<2.5-1.2B-Instruct 🎆♠ and the architecture is actually insane 🚀 🔹 Hybrid design: Convolutions + Attention (not pure transformer) 🔹 1.2B params but rivals 3B+ models 🔹 239 tok/s on AMD CPU (runs under 1GB RAM) 🔹 Built for agentic workflows\ \ JAN 8, 2026](https://www.twitter.com/fahdmirza/status/2009156917860729055)

• [adwaitx.bsky.social\ \

  🔒 Game changer: AMD Ryzen AI runs Liquid AI's LFM2 model locally for meeting summaries. Zero internet needed, complete privacy, 16-second processing. Finally, AI PCs that don't send your data to the cloud. #AdwaitX #AMD #PrivacyFirst #AI #TechNews #CES2026 #AINews #AIComputing #EnterpriseAI\ \ JAN 6, 2026](https://www.adwaitx.com/amd-ryzen-ai-liquid-ai-on-device-meeting-summaries/)

• [Matrix_Memories\ \

  @alshell7 @LiquidAI_ There are also the currently available LFM2-1.2B-Extract and LFM2-350M-Extract. Not tuned for PII but they will do the job very well. Been trying them in the last couple days for a few things and have to say, they are amazing models.\ \ OCT 13, 2025](https://www.twitter.com/Matrix_Memories/status/1977732588916617473)

• [omer_oved\ \

  @maximelabonne Interesting checkpoint. LFM2-2.6B-Exp seems to leverage pure RL effectively for instruction following and reasoning, outperforming most 3B class models in specialized tasks.\ \ DEC 26, 2025](https://www.twitter.com/omer_oved/status/2004744890475819516)

• [IulianHI\ \

  LiquidAILFM2.5-1.2B Review: The best free model for high-speed utilityI’ve been hunting for a model that doesn't feel like a sluggish Transformer for high-frequency, low-latency tasks. I finally spent a few days with LiquidAI’s LFM2.5-1.2B-Instruct, and honestly, the performance profile of this Liquid Neural Network (LNN) architecture is a game changer for edge-style utility. The Use Case I set up a real-time monitor for a cluster of web servers...\ \ JAN 29, 2026](https://www.reddit.com/r/AIToolsPerformance/comments/1qq0v5j/liquidai_lfm2512b_review_the_best_free_model_for/)

• [bluehatone\ \

  AI that runs offline on your phone. LFM2.5-1.2B-Thinking uses under 1GB, CPU only, 10 to 15 tokens per second, about 1.5 second latency. Scores 72% GSM8K, 68% ARC. Apache 2.0 on Hugging Face Jan 20 2026. Faster and lighter than Phi-2 and Gemma 2B, but clouds still lead.\ \ JAN 27, 2026](https://www.twitter.com/bluehatone/status/2016052297789264344)

• [10turtle_com\ \

  @liquidai Impressive results for a 3B model pure RL is paying off.\ \ DEC 26, 2025](https://www.twitter.com/10turtle_com/status/2004510469747568864)

• [ContextrixAi\ \

  @paulabartabajo_ @liquidai This is impressive. A 1.2B model achieving 239 tok/s on an AMD CPU and 82 tok/s on a mobile NPU while using under 1GB RAM makes truly private, offline, production-grade inference on everyday devices realistic now.\ \ JAN 13, 2026](https://www.twitter.com/ContextrixAi/status/2011127756944458117)

• [SwishMoe\ \

  @liquidai is the best model I’ve worked with for post-training and fine-tuning, especially at 2–3B params. Most open-source models (Qwen, DeepSeek, etc.) feel over-tuned and hard to adapt.\ \ JAN 12, 2026](https://www.twitter.com/SwishMoe/status/2010626676103815601)

• [benja0x40\ \

  Liquid AI announced LFM2-VL, fast and lightweight vision models (450M & 1.6B). 2 models based on the hybrid LFM2 architecture: LFM2-VL-450M and LFM2-VL-1.6B. Available quant: 8bit MLX, GGUF Q8 & Q4 (llama.cpp release b6183)\ \ AUG 17, 2025](https://www.reddit.com/r/LocalLLaMA/comments/1msjr8e/liquid_ai_announced_lfm2vl_fast_and_lightweight/)

• [Tech_AI_Tech\ \

  @LiquidAI_ @Shopify Impressive collaboration — sub-20ms inference and measurable conversion gains show how foundational models can deliver real, production-grade impact in commerce. #AI #TechAI\ \ NOV 13, 2025](https://www.twitter.com/Tech_AI_Tech/status/1989032647490998325)

• [stevechen\ \

  @liquidai Incredible performance! Congratulations @liquidai !\ \ DEC 26, 2025](https://www.twitter.com/stevechen/status/2004433825087307859)

• [Kirill Solodskih, Solodskikh\ \

  Good updates coming to TheStage AI On-device SDK.\ \ BTW, new model from Liquid AILFM2.5-230m 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\\ \ JUL 16, 2026](https://www.linkedin.com/feed/update/urn:li:activity:7477319669234802688/)

• [rasmus1610\ \

  LFM2-1.2b is very very usable on my pathetic 8GB M2 MacBook Air. Love it.\ \ JUL 13, 2025](https://www.twitter.com/rasmus1610/status/1944348052564701388)

• [ngxson\ \

  Trying out LFM2 350M from @LiquidAI_ and was mind-blown 🤯 The responses were very coherent. Less hallucinations compared to models of the same size. Very well done!! The best part: Q4_K_M quantization is just 230 Megabytes, wow!\ \ JUL 14, 2025](https://www.twitter.com/ngxson/status/1944819116826640576)

• [guiopen\ \

  Nice! I will test it today, the instruct version punches way above its weight, but I usually don't get good results with small thinking models because they enter in a thinking loop, but it seems there was a focus on preventing that.\ \ Also, there is a mention saying the model is not suitable for coding, do you plan to release a coding capable (even if not code focused) in the future? The previous 8b moe had additional training tokens of code. With the tool call capabilities of lfm + small memory foot print of context length, a code capable lfm2.5 8b moe would be amazing\ \ JAN 20, 2026](https://www.reddit.com/r/LocalLLaMA/comments/1qi512t/liquid_ai_released_the_best_thinking_language/o0pjf9b/)

• [akmessi2810\ \

  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\\ \ JUN 29, 2026](https://www.reddit.com/r/LocalLLaMA/comments/1ugtv27/finetuned_liquidais_lfm25230m_on_fable5_coding/?share_id=JqGGdNoO-cy0QpxQ3LV9S&utm_content=1&utm_medium=ios_app&utm_name=ioscss&utm_source=share&utm_term=1)

• [codewithimanshu\ \

  @MParakhin Wow, Mikhail, that's some impressive performance! Sub-20ms inference is quite something, and the architecture sounds game-changing, right?\ \ NOV 15, 2025](https://www.twitter.com/codewithimanshu/status/1989690758790529163)

• [insignifictcman\ \

  Small but Mighty Liquid AI's LMF2-2.6B-Exp punching much above its weight\ \ DEC 26, 2025](https://www.twitter.com/insignifictcman/status/2004536715734880433)

• [Ealdorwolf\ \

  @LocallyAIApp LFM2 2.6B-Exp-8bit works super fast and efficient in iphone 16 pro\ \ FEB 1, 2026](https://www.twitter.com/Ealdorwolf/status/2018093833251426316)

• [Joshua Lochner\ \

  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\ \ MAR 5, 2026](https://www.linkedin.com/posts/xenova_okay-this-is-actually-insane-you-can-activity-7432494103348539392-EdLC/?utm_source=share&utm_medium=member_desktop&rcm=ACoAABFLa_8B0cHJxxUksizHSKZJablLjsf0cbw)

• [short_circuit32\ \

  @tmikov Right now, among models with <1B params, LFM models are 🔥\ \ FEB 3, 2026](https://www.twitter.com/short_circuit32/status/2018606604426609133)

• [unsolved-problems\ \

  In my limited experience testing various nano/tiny/small models from various labs, LFM2-1.2B was galaxies ahead of competitors. It's 1.2B but really behaves more like a ~10B model in terms of coherence and reasoning power. Even their smaller ~300M model feels more like a 4B, which is mind blowing because it runs fast even on a CPU (hell, it runs fast even on a browser via WebAssembly, unfathomably amazing). I'm hoping that these specialized models will be very good, I have a good amount of trust in Liquid AI.\ \ OCT 1, 2025](https://www.reddit.com/r/LocalLLaMA/comments/1nuyjp9/liquidai_bet_on_small_but_mighty_model/nh8zdjt/)

• [@kryptorina\ \

  🚨 GAME CHANGER ALERT: MIT just broke the AI game! 🤯 Liquid AI's new models are DESTROYING traditional LLMs and here's why this is HUGE 👇✨ Uses 90% FEWER neurons but performs BETTER 🚀 Handles 1 MILLION tokens with minimal memory 💡 Perfect for edge devices (your phone could run this!) 🔥 LFM-1B is setting NEW records on every benchmark. While everyone's obsessing over bigger models, Liquid AI said "hold my beer" and went LIQUID 💧 This isn't just an upgrade - it's a complete paradigm shift. Traditional transformers are about to look like flip phones 📱➡️🧠The future of AI just got a lot more interesting... and efficient! 🌟 What do you think? Are liquid neural networks the future? Drop your thoughts below! 👇#LiquidAI #AI #MachineLearning #TechNews #Innovation #MIT #NeuralNetworks #ArtificialIntelligence #TechBreakthrough #FutureOfAI\ \ OCT 1, 2025](https://www.instagram.com/p/DPQkog6FRhE)

• [hishamkhdair\ \

  Great leap in on-device AI: Liquid's new LFM2.5-1.2B-Thinking does genuine step-by-step reasoning using just ~900 MB RAM, runs on basically any modern phone. Beats larger models like Qwen3-1.7B on math/tool use while being dramatically faster & leaner. Privacy + zero-latency\ \ JAN 21, 2026](https://www.twitter.com/hishamkhdair/status/2014172020838400390)

• [heynerdceo\ \

  @LiquidAI_LiquidAI/LFM2-8B-A1B model is really fast on iPhone. I must say I really like the vibe of this model for an on-device model. Any other LLMs I can check out?\ \ OCT 18, 2025](https://www.twitter.com/heynerdceo/status/1979504508624314513)

• [proudmoontruther.space\ \

  okay no longer trapped inside today :> i started running a shitty vibe coded agent on my jetson orin nano called laine :D (yes from SEL). it’s using LFM2 8B A1B which so far is an amazing model so far. I love it so far\ \ JAN 27, 2026](https://bsky.app/profile/proudmoontruther.space/post/3mdgv3zuh7s2d)

• [romankhrupa\ \

  @pasashque @Apolloaiapp @LiquidAI_ We use LEAP Edge SDK to work with Liquid models. U can check it here ⬇️https://docs.liquid.ai/leap/edge-sdk/android/android-quick-start-guide\\ \ OCT 3, 2025](https://www.twitter.com/romankhrupa/status/1974113120974192724)

• [Mostlime12195\ \

  @maximelabonne @huggingface Congrats! I pre-trained a model using the LFM2 architecture a couple days ago, it's very efficient for training. Forgive me if I'm mistaken, why haven't you guys put out any research directly comparing the LTC architecture with hybrid-attention or regular transformers in LLMs?\ \ JAN 18, 2026](https://www.twitter.com/Mostlime12195/status/2012825023153639640)

• [Marktechpost\ \

  Liquid AI Open-Sources LFM2: A New Generation of Edge LLMs. Liquid AI just dropped a game-changer for edge computing with LFM2, their second-generation foundation models that run directly on your device. These aren't just incremental improvements—we're talking 2x faster inference\ \ JUL 14, 2025](https://www.twitter.com/Marktechpost/status/1944652803412300020)

• [Samunder12or8\ \

  Just tried LFM2 2.6 EXP model and it's mind-blowing to me. It's my go-to model for my phone with ChatterUI 😁 https://huggingface.co/LiquidAI/LFM2-2.6B-Exp\\ \ DEC 26, 2025](https://www.twitter.com/Samunder12or8/status/2004620857432584367)

• [DinaAl_Jamal\ \

  @Prince_Canuma @liquidaiLFM2.5-Audio's end-to-end processing is brilliant-especially those latency cuts! MLX-Audio integration could redefine on-device voice tech. 🚀\ \ JAN 7, 2026](https://www.twitter.com/DinaAl_Jamal/status/2008991689344753739)

• [Temomemo2020\ \

  Stop everything and grab this Vision Language Model fast! 📱👁️ This tiny 3B model is a breakthrough in Edge AI and Computer Vision. It’s faster than the giants and runs perfectly on mobile.  #liquidai #edgeai #computervision #tinymodel #opensource #aitechnology\ \ JAN 3, 2026](https://www.twitter.com/Temomemo2020/status/2007523295504236922)

• [psk90_ai\ \

  🔥 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\ \ FEB 9, 2026](https://www.twitter.com/psk90_ai/status/2020892959596818940)

• [MikeDevPro\ \

  LFM2 GGUF models just dropped for llama.cpp - these are game changers. 2x faster than Qwen3 on CPU, 200% higher throughput vs competitors, designed for edge deployment. Finally, high-performance models that actually run well locally: https://huggingface.co/LiquidAI/LFM2-1.2B-GGUF\\ \ JUL 13, 2025](https://www.twitter.com/MikeDevPro/status/1944331198685208654)

• [Emp0_com\ \

  LFM2-2.6B-Exp shows pure RL beating giants: a 3B model tops a 263x larger peer on IFBench, with think tokens and on device tool use for edge reasoning. Is edge AI finally reality? #EdgeAI\ \ DEC 28, 2025](https://www.twitter.com/Emp0_com/status/2005256928335708556)

• [iMATTHEWRYAN\ \

  @liquidai Bottom line:LFM2 models show that you can get 3B–4B-level quality from ~1B–2.6B models and run them 2–3× faster on a phone. Small models, if designed correctly, can dominate the accuracy–latency tradeoff.\ \ DEC 4, 2025](https://www.twitter.com/iMATTHEWRYAN/status/1996586695224070489)

• [varvapally\ \

  @its_maddy_a @liquidai @ZeroGPU_AI This is awesome to hear! 🚀 Liquid AI’s architecture is a game-changer for efficiency. We’d love to see how @ZeroGPU_AI pushes the limits of the LEAP SDK. Keep us posted on the progress!\ \ JAN 8, 2026](https://www.twitter.com/varvapally/status/2009357303598797201)

• [AxlysCustoms\ \

  @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)\ \ FEB 15, 2026](https://www.twitter.com/AxlysCustoms/status/2023193527660802269)

• [Xenova@xenovacom\ \

  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.\ \ JUN 25, 2026](https://x.com/xenovacom/status/2070210622239707568)

• [support_huihui\ \

  @AntDX316 The response quality of LiquidAI/LFM2.5-1.2B-Thinking and LFM2.5-1.2B-Instruct is very high, which is quite remarkable for such small models.\ \ JAN 25, 2026](https://www.twitter.com/support_huihui/status/2015346504941318551)

• [singleapi2\ \

  1. Exciting AI breakthroughs are here! New compact models like LFM2.5 bring powerful on-device intelligence with billions of parameters and smarter multimodal abilities. Tencent’s Youtu-LLM outperforms bigger models in coding & math tasks. 🤖✨\ \ JAN 6, 2026](https://www.twitter.com/singleapi2/status/2008487114355114434)

• [BearleDev78579\ \

  Game-changing on-device AI just dropped. 🤯Liquid AI unleashed LFM2-Audio, a 1.5B param speech model with a crazy low 95ms latency—that's literally the speed of a blink. This is huge. Try it here: https://playground.liquid.ai/login?callbackUrl=%2FtalkBut\\ that's not all... 🧵\ \ OCT 1, 2025](https://www.twitter.com/BearleDev78579/status/1973414185154367566)

• [the_ai_scope\ \

  🚀 LiquidAI just released LFM2‑2.6B‑Exp, a 3 B‑parameter text‑gen model that scored 683 on the latest benchmark — updated just 3 hrs ago. It brings near‑GPT‑4 quality with a tiny footprint. https://huggingface.co/LiquidAI/LFM2-2.6B-Exp\\ \ DEC 26, 2025](https://www.twitter.com/the_ai_scope/status/2004508756206321693)

• [yijiezuimei\ \

  @LisaFlorentina8 Liquid AI is the future for sure\ \ SEP 13, 2025](https://www.twitter.com/yijiezuimei/status/1966895022009352311)

• [simpaisush\ \

  For ~3B model size, its instruction-following and reasoning results are genuinely impressive.\ \ DEC 25, 2025](https://www.twitter.com/simpaisush/status/2004223889929637975)

• [Archit_AK47\ \

  The rise of Liquid AI is worth watching 👀. Built for efficiency, scalability & speed — it’s redefining how AI systems learn and adapt. Feels like we’re just scratching the surface of what this tech can unlock. 🚀 #AI #LiquidAI #DeepTech\ \ SEP 26, 2025](https://www.twitter.com/Archit_AK47/status/1971647610940543225)

• [adityaberry2004\ \

  @TheAhmadOsman These are big names the real benificial models are the ones that are bringing the cost of intelehence down namely the 1-10 b prams pioneerd by the likes of @liquidai and @GeminiApp with gemma with their amazing 1 b prams model which can run inside your mobile phones natively\ \ JAN 26, 2026](https://www.twitter.com/adityaberry2004/status/2015711903256047986)

• [bygregorr\ \

  @paulabartabajo_ @liquidai Running models locally can be a game changer for privacy and cost. It's amazing how far we've come with efficient inference. What are the implications for offline apps now?\ \ JAN 29, 2026](https://www.twitter.com/bygregorr/status/2016880416683114958)

• [MartinSzerment\ \

  Small models are getting scary good. Liquid AI shows how reinforcement learning can turn a 2.6B model into something that beats most 3B LLMs — perfect for edge, local and real-world deployments. #AI #LLM #EdgeAI #MachineLearning #ReinforcementLearning\ \ DEC 26, 2025](https://www.twitter.com/MartinSzerment/status/2004587273258692657)

• [HuggingModels\ \

  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.\ \ FEB 17, 2026](https://www.twitter.com/HuggingModels/status/2023939149548646730)

• [Noctusnoctus91\ \

  One thing I like about @opencode is the variety of models on the Go plan, but deciding which one to use every time gets annoying.\ \ So I spent some time experimenting with @liquidaiLFM2.5-Encoder-350M-Prompt-Router model.\ \ I know routing between LLMs isn't really the model's intended use case, but its zero-shot routing head made for a fun experiment. A single encoder pass (~250ms) classifies the prompt and routes it to one of the 12 models available in OpenCode Go plan.\ \ Still early, but it's been a fun way to get more out of the Go plan.\ \ JUL 28, 2026](https://x.com/noctus91/status/2082160008402522341?s=20)

• [tikooww\ \

  @localghost Big congrats, on device and Liquid AI feels like a game changer combo.\ \ JUL 22, 2025](https://www.twitter.com/tikooww/status/1947781969737126327)

• [Thrumpwart\ \

  Very good little model released quietly. In testing it's quite competent and very fast. Quants available on HF.\ \ SEP 24, 2025](https://www.reddit.com/r/LocalLLaMA/comments/1nprim2/introducing_lfm226b_redefining_efficiency_in/ng1csp2/)

• [YigitMertCahit\ \

  @maximelabonne @Lightricks Seeing Liquid AI models trending is a strong signal for non-Transformer architectures! 💧 At 1.2B params, these are perfect for experimenting with efficient local inference. Great to see diversity in the open-source leaderboard! 🚀 #AI #HuggingFace\ \ JAN 11, 2026](https://www.twitter.com/YigitMertCahit/status/2010396372923826237)

• [eustachelb\ \

  Cool release by @LiquidAI_: LFM2-Audio-1.5B It’s a pretty cool omni-architecture that enables prediction of both text and audio tokens, meaning it can handle multi-turn S2S, ASR, and TTS (with voice description) within a single model. Great to see, once again this year, a model\ \ OCT 3, 2025](https://www.twitter.com/eustachelb/status/1974129698960453847)

• [gm8xx8\ \

  🔥 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..\ \ AUG 12, 2025](https://www.twitter.com/gm8xx8/status/1955336424837857392)

• [PriyaonAI\ \

  LiquidAI is unleashing new real-time translation power! 🤯 Introducing LFM2-350M, a model fine-tuned for Japanese-English communication! This is a total game-changer for breaking language barriers! 🚀✨\ \ SEP 8, 2025](https://www.twitter.com/PriyaonAI/status/1964922507712180438)

• [LaurencePostrv\ \

  Hell yes! Currently running a 2.6B liquid foundation model on my phone. Imagine how useful this could be in any number of scenarios without network access!\ \ OCT 2, 2025](https://www.twitter.com/LaurencePostrv/status/1973897139328856399)

• [Cubow\ \

  It is incredible, hard to believe a 1B model can sound coherent, but you can straight up have conversations with it. Got me excited for Gemma4 and the future of smaller models\ \ JAN 16, 2026](https://www.reddit.com/r/LocalLLaMA/comments/1qdax6z/lfm_25_is_insanely_good/nzvpt45/)

• [SimslearnAi\ \

  @liquidai Whoa, a 3B model outperforming one 263× its size? This is next-level RL magic!\ \ ‍\ \ DEC 25, 2025](https://www.twitter.com/SimslearnAi/status/2004292056076988709)

• [DittmannAxel\ \

  450M vision model runs at 353 tokens/sec on MLX using just 1.456GB memory 🔥. LiquidAI drops LFM2-VL-450M → MLX-VLM community with @Prince_Canuma ports it → Your Mac becomes a vision AI beast. Liquid neurons adapting computation on the fly. 450M params punching way above its weight class. Open source wins again. Model drops, community delivers, boom - it's local. What are you building with instant vision AI?\ \ AUG 18, 2025](https://www.twitter.com/DittmannAxel/status/1957360463764746471)

• [kento_morota\ \

  📰 LFM2 2.6B-Exp on Android: 40+ TPS and 32K context. The article discusses the impressive performance of LiquidAI's new LFM2 2.6B-Exp model, which can achieve GPT-4 level performance across a wide range of benchm… Read more 👇https://ai-curator.jp/articles/cmjw8efiq00v2ct1epceqo9q1 #AI #AINews\ \ JAN 1, 2026](https://www.twitter.com/kento_morota/status/2006924073016897591)

• [mojombo\ \

  Two amazing companies, now more amazing with this partnership!You're going to start hearing a lot more about @LiquidAI_ (a @PWVentures investment) in the future; their novel AI architecture blows away transformers in a ton of valuable use cases.Congrats to both sides on\ \ NOV 13, 2025](https://www.twitter.com/mojombo/status/1989045577796268332)

• [Samuel5am2x\ \

  Jev-inspired inference. 350M parameters. 63× faster !\ \ Spent a few hours trying Jev-style inference with LFM2.5-350M.\ \ 63× faster on an L40S. 8x on MPS. No training (for now) just parallel decisions.\ \ SEP 16, 2026](https://x.com/5am2x/status/2100157777595314615?s=20)

• [Russet-Mod\ \

  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.\ \ FEB 8, 2026](https://www.reddit.com/r/Russet/comments/1qzj5gz/benchmarking_ondevice_mlx_llms_with_russet_on/)

• [Daniel van Strien@vanstriendaniel\ \

  @liquidai AI just released LFM2.5 encoders small bidirectional models with 8k context.\ \ I've added a training recipe for fine-tuning them on @huggingface Jobs (works for any encoder), and used it to train a classifier that suggests task categories for Hub datasets from their READMEs.\ \ One command, 20 minutes on an A100, about $0.80.\ \ https://danielvanstrien.xyz/posts/2026/agent-trained-classifier/\\ \ JUL 29, 2026](https://x.com/vanstriendaniel/status/2082444493467382103)

• [foundersignals\ \

  Liquid AI released an experimental checkpoint of LFM2-2.6B using pure RL, making it the strongest 3B on the market. Meet the strongest 3B model on the market. LFM2-2.6B-Exp is an experimental checkpoint built on LFM2-2.6B using pure reinforcement learning. Consistent\ \ DEC 27, 2025](https://www.twitter.com/foundersignals/status/2005031183001026590)

• [steezy13312\ \

  I got frustrated dealing with massive responses from many MCPs and threw something together over the last couple days... it might help you too. Or not!...you're playing with multiple AI tools/coding assistants and hate having to reconfigure MCPs for each one- Very configurable options to override behavior globally or different tools via a single JSON file, plus a UI for management and visibility. I've been testing with a high-quant Qwen3-0.6b and LFM2-1.2b and it's doing very well for me. For example, I have it use web search and fetch for URLs and instead of having the larger model process the entire pages, the tiny model reads the page up to 10x faster, and just gives the large model the answers it needs, also keeping context lower.\ \ JAN 2, 2026](https://www.reddit.com/r/LocalLLaMA/comments/1q2et07/i_got_frustrated_dealing_with_massive_responses/)

• [brooksy4503\ \

  @kimmonismus I tried this on my phone and it's a really good model. I enjoyed using it.\ \ JAN 7, 2026](https://www.twitter.com/brooksy4503/status/2008822236871065641)

• [fragermk\ \

  Local LLM models are getting good. LFM2-2.6B-Exp is quite powerful and fast on Apple Silicon.\ \ DEC 25, 2025](https://www.twitter.com/fragermk/status/2004378374492577849)

• [KaroYadgar\ \

  Liquid AI released an experimental checkpoint of LFM2-2.6B using pure RL, making it the strongest 3B on the market. "Meet the strongest 3B model on the market. LFM2-2.6B-Exp is an experimental checkpoint built on LFM2-2.6B using pure reinforcement learning. Consistent improvements in instruction following, knowledge, and math benchmarks. Outperforms other 3B models in these domains. Its IFBench score surpasses DeepSeek R1-0528, a model 263x larger"\ \ DEC 26, 2025](https://www.reddit.com/r/singularity/comments/1pvypgn/liquid_ai_released_an_experimental_checkpoint_of/)

• [DoDataThings\ \

  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 👊🏾\ \ FEB 2, 2026](https://www.twitter.com/DoDataThings/status/2018377798789046517)

• [kimmonismus\ \

  Holy moly, Liquid AI just unveiled LFM2.5, a powerful open-weight model family designed to run fast, private, and always-on directly on devices. LFM2.5 sets a new bar for edge AI across text, vision, audio, and Japanese-language use cases.The TTS examples are amazing btw\ \ JAN 6, 2026](https://www.twitter.com/kimmonismus/status/2008531535264469028)

• [TeksEdge\ \

  🤯 🎁 Liquid AI brings "Deep Thinking" to a 2.6B model / your Phone / your robot 🤖Liquid AI released a holiday present for your new robot 🎁 They just cracked the code on efficient intelligence. With LFM 2-2.6B-Exp, the era of bloated, memory-hogging Transformers running on\ \ DEC 25, 2025](https://www.twitter.com/TeksEdge/status/2004298566055997493)

• [birdman1710\ \

  @songdng @liquidai Just tried it peak memory usage of barely around 2.5Gb for a thinking model is crazy good\ \ JAN 21, 2026](https://www.twitter.com/birdman1710/status/2013915451982807466)

• [Crypto_neowolf\ \

  Amazed by LFM 3b model.. and amazed by the @liquidai team for making this amazing model within 3 b parameter... running quantized Q4 but already loving it... earlier i was running gemma2 but looks like LFM is my new love...\ \ DEC 26, 2025](https://www.twitter.com/Crypto_neowolf/status/2004611371900829904)

• [Foreign-Beginning-49\ \

  Damn you guys keep knocking it out of the park. Congratulations on all the amazing releases. Im still trying to fully utilize LFM2-1.2B then you drop new weights. So fun to be here and experience the fruits of your toil. Thanks for looking out for those of us with less computational capacities. Its amazing how far small models have come and you're an integral part of that puzzle. Best wishes\ \ JAN 7, 2026](https://www.reddit.com/r/LocalLLaMA/comments/1q6nm6a/liquid_ai_releases_lfm226btranscript_an/ny91glc/)

• [red_bear888\ \

  Hermes 4 14B: Completely uncensored and answers all the questions that other LLMs reject. Jan-Nano: Excellent agentic model for using tools and automating many tasks LFM2-VL 1.6B: Light weight and extremely fast multimodal model with vision\ \ OCT 11, 2025](https://www.twitter.com/red_bear888/status/1976926992961544357)

• [prithvii_J\ \

  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\ \ FEB 13, 2026](https://www.twitter.com/prithvii_J/status/2022360679542591597)

• [KokaOP\ \

  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..\ \ FEB 14, 2026](https://www.reddit.com/r/LocalLLaMA/comments/1r4e3w3/kanitts2_our_texttospeech_model_with_framelevel/)

• [biXente_Latte\ \

  @paulabartabajo_ @liquidai I love this model. It makes of Llama 3.2 back in the time. Inference speed is also insane on Ollama and Apollo. I want to try it with llama.cpp cli. Liquid AI is choosing a really promising path in AI and you are pushing the boundaries of SLM\ \ JAN 29, 2026](https://www.twitter.com/biXente_Latte/status/2016869443029147700)

• [sorbusCobPhiil\ \

  @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. 🤌\ \ FEB 5, 2026](https://www.twitter.com/sorbusCobPhiil/status/2019608956164067689)

• [alok_nayak\ \

  #LiquidAI, an @MIT spinoff, just released a blueprint for enterprise-grade small models. Small can be mighty: high performance, low latency, on-device #AI. 51-page roadmap now public. 🚀#EnterpriseAI #OnDeviceAI @VentureBeat @carlfranzen @liquidai\

\ DEC 2, 2025](https://www.twitter.com/alok_nayak/status/1995795470174842956)

• [jokerbobx\ \

  5️⃣ Jan Nano An underrated agent model. Executes, plans, automates. Feels like a real assistant. 6️⃣ LFM2 VL 1.6BCompact multimodal engine. Fast inference. Handy for on device vision tasks. 7️⃣ Magistral Small 1.2 Solid visual plus language mix. Consistent and easy to use.\ \ OCT 17, 2025](https://www.twitter.com/jokerbobx/status/1979104487697715379)

• [RahulVerma989\ \

  @liquidai Impressive benchmarks for LFM2.5-1.2B-Thinking, especially with on-device capabilities and concise reasoning - a significant step forward in AI efficiency.\ \ JAN 21, 2026](https://www.twitter.com/RahulVerma989/status/2013850018999161193)

• [ShubhamMal72313\ \

  What if BabyAGI didn’t need the cloud? We took @yoheinakajima's BabyAGI loop and ran it locally on an iPhone over a small language model (LFM2-350M). No OpenAI key. No servers. Just @RunAnywhereAI SDK + @liquidaiLFM2-350M running everything fully on-device.\ \ NOV 28, 2025](https://www.twitter.com/ShubhamMal72313/status/1994508784304300376)

• [xeophon\ \

  @liquidai Qwen3 4B 2507 Instruct might be an unfair comparison cause it is 2x the size, but thats what I daily drive. When I compare LFM against Qwen3 1.7B or Granite, LFM absolutely crushes the others and it is not even close. LFM2.5 is the best <4B model currently out there.\ \ JAN 6, 2026](https://www.twitter.com/xeophon/status/2008444769874100575)

Show more