BTC $84,714.03 +0.25%
ETH $2,685.06 +0.76%
BNB $785.59 +2.56%
XRP $1.49 +0.72%
SOL $119.64 +1.25%
TRX $0.3352 +0.30%
DOGE $0.0929 +0.92%
ADA $0.2448 +1.55%
BCH $316.47 +2.84%
LINK $14.04 +2.65%
HYPE $89.48 +3.71%
AAVE $180.52 -0.27%
SUI $1.17 +4.34%
XLM $0.2157 +1.16%
ZEC $1,316.04 +1.92%
AAPL $333.27 -0.02%
AMZN $252.01 +0.41%
GOOGL $343.23 -0.00%
MSFT $518.03 +0.09%
META $730.14 +0.37%
NVDA $234.94 +0.44%
TSLA $371.00 -0.06%
SNDK $1,718.67 -0.04%
INTC $117.96 -1.16%
SPCX $159.03 +0.13%
MU $1,070.34 -0.15%
AMD $634.04 +0.16%
BTC $84,714.03 +0.25%
ETH $2,685.06 +0.76%
BNB $785.59 +2.56%
XRP $1.49 +0.72%
SOL $119.64 +1.25%
TRX $0.3352 +0.30%
DOGE $0.0929 +0.92%
ADA $0.2448 +1.55%
BCH $316.47 +2.84%
LINK $14.04 +2.65%
HYPE $89.48 +3.71%
AAVE $180.52 -0.27%
SUI $1.17 +4.34%
XLM $0.2157 +1.16%
ZEC $1,316.04 +1.92%
AAPL $333.27 -0.02%
AMZN $252.01 +0.41%
GOOGL $343.23 -0.00%
MSFT $518.03 +0.09%
META $730.14 +0.37%
NVDA $234.94 +0.44%
TSLA $371.00 -0.06%
SNDK $1,718.67 -0.04%
INTC $117.96 -1.16%
SPCX $159.03 +0.13%
MU $1,070.34 -0.15%
AMD $634.04 +0.16%

seed

全部
文章
快訊

first_img 字節討論訓練超 5 萬億參數模型,Seed 基模團隊調整架構

ChainCatcher 消息,据晚点 LatePost 報導,字節跳動正在討論訓練參數規模超 5 萬億的大模型,超過阿里 Qwen 3.8-Max(2.4 萬億)和月之暗面 K3(2.8 萬億),為目前國內已知最大規模方案。該計劃仍處早期,不代表最終一定發布。新模型擬由 Seed Foundation 負責人項亮主導,並與大語言模型預訓練數據負責人沈科合作,Seed 正據此重新梳理組織、劃分職責並分配資源。兩週前,字節跳動創始人張一鳴與 Seed 負責人吳永輝召開全員會。張一鳴安撫團隊稱訓大模型本就很難,一段時間內可接受落後,希望以追求智能上限為目標、跻身世界第一梯隊;他認可編程是當前關鍵方向,主張整合火山引擎、飛書與豆包資源以構築算力和數據優勢,同時提醒勿被單一熱點牽著走。他表揚 Seedance 的差異化領先,並明確反對蒸餾,認為其難以真正超越,應從更底層構建 AGI 壁壘,並表示公司將在 AI 上持續加大投入。過去半年 Seed 多模態表現突出,Seedance 2.0、Seedream 等帶動火山引擎 MaaS,但語言模型 Seed 2.0 市場反響有限,Coding 能力落後影響收入結構。字節已高薪引入郭達雅專攻 Coding,並收攏相關資源。面對行業普遍做大模型尺寸的趨勢,字節希望以更大規模爭取彎道超車,同時推動取消賽馬、打破部門牆,集中力量攻堅。
app_icon
ChainCatcher 與創新者共建Web3世界