{"rewrite":{"id":"r_d702b717387db38e6db6cf1e","clusterId":"c_7a0122a5ee5f9003fc9bd96c","slug":"liquid-ai-releases-lfm3-suite-of-six-open-models-for-phones-and-laptops","model":"deepseek-v4-1-flash","headline":"Liquid AI Releases LFM3 Suite of Six Open Models for Phones and Laptops","summary":"Liquid AI released LFM3, a family of small AI models that run on smartphones and laptops. The six models derive from LFM3-1.7B-Base, pretrained on 22 trillion tokens of scientific papers, code, and math problems. Liquid AI developed the suite with AMD and Qualcomm, and CEO Ramin Hasani promoted it alongside AMD CEO Lisa Su at a January 2026 CES keynote. Every model, including the base, is open. LFM3-1.7B decodes at 43.6 per second on an iPhone 17 Pro with 1.90GB of memory use.","whyItMatters":"Liquid AI is staking its case on device memory and open weights rather than raw scale: LFM3-VL-2B posts higher benchmark scores than Qwen3.5 2B and Gemma 4 E2B, and in some tests beats the larger Gemma 4 12B, while the 1.7B model fits in 1.90GB on a phone.","webCardHtml":"\u003cp\u003eLFM3-VL-2B, the vision-language member, records higher benchmark scores than Qwen3.5 2B and Gemma 4 E2B, and in some tests surpasses the much larger Gemma 4 12B. That comparison is the argument behind the release: a model small enough for a laptop or a phone need not concede accuracy to others of comparable scale. The two d1 models step away from chat altogether and return judgments on options with confidence scores, a format built for decision tasks.\u003c/p\u003e\u003cul\u003e\u003cli\u003e\u003cstrong\u003eLFM3-1.7B\u003c/strong\u003e: language model with text input and output\u003c/li\u003e\u003cli\u003e\u003cstrong\u003eLFM3-VL-2B\u003c/strong\u003e: vision-language model taking text, image, and video input\u003c/li\u003e\u003cli\u003e\u003cstrong\u003eLFM3-Audio-2B\u003c/strong\u003e: audio model for text and audio input and output\u003c/li\u003e\u003cli\u003e\u003cstrong\u003eLFM3-Omni-2B\u003c/strong\u003e: omni model taking text, image, video, and audio, returning text and audio\u003c/li\u003e\u003cli\u003e\u003cstrong\u003ed1-2B\u003c/strong\u003e: decision model for text, images, and video\u003c/li\u003e\u003cli\u003e\u003cstrong\u003ed1-Omni-2B\u003c/strong\u003e: decision model adding audio to text, images, and video\u003c/li\u003e\u003c/ul\u003e","blueskyPost":"Liquid AI built LFM3 with AMD and Qualcomm, and CEO Ramin Hasani showed it beside AMD CEO Lisa Su at a January 2026 CES keynote. The 1.90GB iPhone footprint is the number those two chipmakers care about.","twitterPost":"LFM3's 1.90GB footprint on an iPhone 17 Pro is the number AMD and Qualcomm signed up for.","threadsPost":"Liquid AI's LFM3 ran on an iPhone 17 Pro at 43.6 tokens per second using 1.90GB of memory. That memory figure is why AMD and Qualcomm are attached to this suite at all: small enough models change which hardware can run them.","newsletterBlurb":"Liquid AI released LFM3, a six-model suite derived from a 1.7B base pretrained on 22 trillion tokens, with every model open. The company built it with AMD and Qualcomm and put CEO Ramin Hasani on stage with AMD CEO Lisa Su at a January 2026 CES keynote. Two d1 decision models return judgments on options with confidence scores instead of conversational text.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20261029-liquid-ai-lfm3/\",\"title\":\"LFM3, a suite of small, high-performance AI models that run on smartphones, arrives: next-generation models including language, vision-language, audio, omni, and decision models\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":6295,"outputTokens":3435,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1793263800,"createdAt":"2026-10-29T08:45:15.000Z","publishedAt":"2026-10-29T08:48:25.000Z","updatedAt":"2026-10-29T08:48:25.000Z"},"cluster":{"id":"c_7a0122a5ee5f9003fc9bd96c","canonicalTitle":"スマホで動作する小型高性能なAIモデル群「LFM3」が登場、言語モデル・視覚言語モデル・音声モデル・オムニモデル・意思決定モデルを含む次世代モデル","representativeArticleId":"a_ae085a05b11cb60097bea83b","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[],\"studios\":[],\"people\":[],\"type\":\"news\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-10-29T08:00:00.000Z","lastSeenAt":"2026-10-29T08:00:00.000Z","updatedAt":"2026-10-29T08:48:25.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20261029-liquid-ai-lfm3/","title":"スマホで動作する小型高性能なAIモデル群「LFM3」が登場、言語モデル・視覚言語モデル・音声モデル・オムニモデル・意思決定モデルを含む次世代モデル"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":[],"studios":[],"people":[],"type":"news","domain":"other","is_roundup":false},"keyFacts":["Liquid AI released LFM3, a six-model suite that runs on smartphones and laptops, with every model open.","LFM3-1.7B-Base was pretrained on 22 trillion tokens covering scientific papers, code, and math problems.","Liquid AI CEO Ramin Hasani appeared with AMD CEO Lisa Su at a January 2026 CES keynote to promote LFM3.","LFM3-1.7B decodes at 43.6 per second on an iPhone 17 Pro with 1.90GB of memory use.","LFM3-VL-2B records higher benchmark scores than Qwen3.5 2B and Gemma 4 E2B, and surpasses Gemma 4 12B in some tests."],"pageLinks":{"healthSlugs":[],"storySlugs":[],"entities":[],"stories":[],"entitiesOk":true,"storiesOk":true}}