{"rewrite":{"id":"r_6cf7b8f04ce2eefc549f16c5","clusterId":"c_75e5f0b114dd5611ef22ee4d","slug":"liquid-ai-applies-dspark-to-lfm2-5-more-than-doubling-speed-on-h100-and-macbook","model":"deepseek-v4-flash:free","headline":"Liquid AI Applies DSpark to LFM2.5, More Than Doubling Speed on H100 and MacBook","summary":"Liquid AI applied the speculative decoding technology DSpark to three models in its compact LFM2.5 series: LFM2.5-1.2B-Instruct, LFM2.5-2.6B, and LFM2.5-8B-A1B. The DSpark-applied models achieved more than a 2x speedup while maintaining performance. LFM2.5-2.6B reached a 2.67x speedup on H100 and 2.27x on a MacBook Pro with M4 Max.","whyItMatters":"The DSpark-applied LFM2.5 models show speculative decoding can push compact on-device models past a 2x speedup without trading accuracy, which Liquid AI frames as a step toward running Fable-level intelligence on smartphones.","webCardHtml":"\u003cp\u003eLiquid AI built separate draft models for each of the three DSpark versions. The draft model for LFM2.5-1.2B-Instruct has 295.7 million parameters, while LFM2.5-2.6B and LFM2.5-8B-A1B each use a 327.7 million parameter draft model.\u003c/p\u003e\u003cp\u003eOn the MacBook Pro with M4 Max, LFM2.5-1.2B-Instruct achieved a 2.54x speedup and LFM2.5-8B-A1B reached 1.18x. The DSpark-applied models are distributed on Hugging Face under the LiquidAI organization.\u003c/p\u003e\u003cp\u003eLiquid AI\u0026#39;s Piotr Mazurek called speculative decoding a key part of compressing Fable-level intelligence to run on smartphones, describing this result as a first step toward future performance gains.\u003c/p\u003e\u003cul\u003e\u003cli\u003e\u003cstrong\u003eLFM2.5-1.2B-Instruct-DSpark\u003c/strong\u003e: 2.54x speedup on MacBook Pro with M4 Max\u003c/li\u003e\u003cli\u003e\u003cstrong\u003eLFM2.5-2.6B-DSpark\u003c/strong\u003e: 2.67x speedup on H100 and 2.27x on MacBook Pro with M4 Max\u003c/li\u003e\u003cli\u003e\u003cstrong\u003eLFM2.5-8B-A1B-DSpark\u003c/strong\u003e: 1.18x speedup on MacBook Pro with M4 Max\u003c/li\u003e\u003c/ul\u003e","blueskyPost":"Liquid AI's DSpark speeds up LFM2.5 by over 2x without a performance hit, making speculative decoding practical on consumer hardware like a MacBook Pro with M4 Max.","twitterPost":"Liquid AI's DSpark doubles LFM2.5 speed on both H100 and MacBook Pro, suggesting speculative decoding is now viable outside data centers.","threadsPost":"Liquid AI's DSpark brings speculative decoding to compact models, more than doubling LFM2.5 speed on both H100 and a MacBook Pro with M4 Max. Running 2.6B at 2.67x on a server GPU and 2.27x on a laptop hints the technique is no longer confined to data centers.","newsletterBlurb":"Liquid AI applied the DSpark speculative decoding technology to three compact LFM2.5 models, more than doubling generation speed while maintaining accuracy. The 2.6B model reached 2.67x on H100 and 2.27x on a MacBook Pro with M4 Max. All three DSpark versions are now distributed on Hugging Face.","attributionJson":"[{\"source\":\"GIGAZINE\",\"url\":\"https://gigazine.net/news/20260821-lfm2-5-dspark-faster-inference/\",\"title\":\"DSpark-Applied Version of Compact Model 'LFM2.5' Appears, More Than Doubling Speed with Speculative Decoding Technology\"}]","lintFlagsJson":null,"lintHits":0,"costUsd":0,"inputTokens":4876,"outputTokens":882,"status":"published","repairAttempts":0,"nextRepairAt":null,"factsAttemptedAt":1787349004,"createdAt":"2026-08-21T21:44:40.000Z","publishedAt":"2026-08-21T21:46:44.000Z","updatedAt":"2026-08-21T21:44:40.000Z"},"cluster":{"id":"c_75e5f0b114dd5611ef22ee4d","canonicalTitle":"小型モデル「LFM2.5」を2倍以上高速化する投機的デコーディング技術DSpark適用版が登場","representativeArticleId":"a_08c9a999848b53e7f2c28edb","sourceCount":1,"writtenSourceCount":1,"writeAttempts":0,"isSolo":true,"entitiesJson":"{\"anime_titles\":[],\"manga_titles\":[],\"work_titles\":[\"LFM2.5\",\"DSpark\"],\"studios\":[\"Liquid AI\"],\"people\":[],\"type\":\"news\",\"domain\":\"other\",\"is_roundup\":false}","contentType":"news","status":"published","firstSeenAt":"2026-08-21T09:14:00.000Z","lastSeenAt":"2026-08-21T09:14:00.000Z","updatedAt":"2026-08-21T21:46:45.000Z"},"attribution":[{"source":"GIGAZINE","url":"https://gigazine.net/news/20260821-lfm2-5-dspark-faster-inference/","title":"小型モデル「LFM2.5」を2倍以上高速化する投機的デコーディング技術DSpark適用版が登場"}],"entities":{"anime_titles":[],"manga_titles":[],"work_titles":["LFM2.5","DSpark"],"studios":["Liquid AI"],"people":[],"type":"news","domain":"other","is_roundup":false},"keyFacts":null}