Liquid AI發表非Transformer的AI模型
支付動態 · 2024-10-01

由4名MIT研究人員聯合創辦的Liquid AI,利用非Transformer架構所建置的LFM基礎模型家族,宣稱在基準測試表現上,已凌駕其他相同規模的Transformer模型,包括Meta Llama、微軟 Phi、Stability Stable以及Google Gemma

/Liquid AI

其中,LFM-1.3B在許多基準測試中擊敗了1B領域的眾多好手,包括蘋果的OpenELM、Meta的Llama 3.2、微軟的Phi 1.5或是Stability的Stable LM 2,Liquid AI說,這是首次有非GPT架構的模型明顯超越Transformer模型。

而LFM-3.1B不僅超越了3B領域的各種Transformer模型、混合模型與RNN模型,甚至超越了前一代的7B與13B模型,不管是Google的Gemma 2,蘋果的AFM Edge AFM,Meta的Llama 3.2,微軟的Phi-3.5都屈居下風。

LFM-40.3B則強調模型規模與輸出品質之間的平衡,儘管擁有400億個參數,但執行時啟用120億個參數來推論,能比美比自己還大的模型,此外,其MoE架構能實現更高的吞吐量,也能部署在更具成本效益的硬體上。

現階段的LFMs擅長一般知識及專業知識,也能處理數學及邏輯推理,並可有效率地處理長文任務,主要支援英文,也支援中文、法文、德文、西班牙文、日文、韓文與阿拉伯文。較不擅長的則有零樣本的程式碼任務,精確的數值計算,具時效性的資訊,也無法計算Strawberry這個字中有多少個R,以及尚未部署人類偏好的優化。

熱門文章
Gaming & Technology Expo Makes a Powerful Entrance in CDMX
Marketing
British gambling levy rates confirmed for each vertical
Regulation
Manila delivers: Highlights from SiGMA Asia 2026 
Southeast Asia
New Jersey July Gambling Revenue Hits $606M, Sweeps Casinos Banned
Regulation
Institutional Academy that exceeded expectations marked the opening of GAT CDMX
Online Game
B2B Tech Infrastructure Gains Momentum in Philippine Gaming Sector
Southeast Asia
That’s a Wrap: AffPapa Conference Madrid 2026 Highlights
Online Game
JILI Partners with Cricket Legend AB de Villiers (ABD) to Launch Exclusive Branded Game Series 100% 11
Sports Game
Indiana online casino bill stalls in House committee
Regulation
Zenith partners with HUIDU for 2026 World Cup Carnival Official Tour
Online Game
Full House at GAT Expo Cartagena 2026 Academic Agenda
Online Game
SBC Summit Canada to Make Player Safety a Key Pillar of 2026 Agenda
Marketing
Kazakhstan plans to penalise online casino promotions
Regulation
PropellerAds Positions Itself as a Go-To Traffic Source for iGaming Advertisers Ahead of a High-Demand Season
Marketing
DB Gaming Is One of Asia’s Top-Tier Game Solution Providers, Recognized as a Pioneer in the Asian Gaming Industry.
Marketing
首頁
遊戲
合作
發現
我的