Several major models shipped in the last 48 hours. ððŒð¿ ð¥ðð ð®ð»ð± ð®ðŽð²ð»ð ð¯ðð¶ð¹ð±ð²ð¿ð,...
- DeepSeek V4éåé«å®¹éãææ¬ææçRAGåºçšã
- GPT-5.5åšé¡¶çº§æšçåå·¥å ·äœ¿çšæ¹é¢è¡šç°æ䜳ã
- Qwen3.6æéåç§æææ¬å°éšçœ²ã
ç»ææ纲
AI æ¿äœ 读äžéåæŽçåºçæ žå¿å±çº§ã
æ绎富åŸ
çšäžåŒ åŸçæž äž»é¢ä¹éŽçå ³ç³»ã
æ¥ç倧纲ææ¬ïŒæ éç¢ / æ JS å奜ïŒ
- Qwen3.6
éå¥ / Highlights
åŒåŸæ¶èäžå享çå ³é®å¥ã
DeepSeek V4 is best for high-volume, cost-sensitive RAG.
â 第 2 段
äžèœœéå¥å¡ PNGGPT-5.5 is best for agents that need top-tier reasoning and tool use.
â 第 3 段
äžèœœéå¥å¡ PNGQwen3.6 is best for private or on-premise deploys.
â 第 4 段
äžèœœéå¥å¡ PNG
ðð²ð²ðœðŠð²ð²ðž ð©ð° ð¶ð ð¯ð²ðð https://t.co/uFAZx3nOcr" / X
Post
Conversation
Several major models shipped in the last 48 hours. ððŒð¿ ð¥ðð ð®ð»ð± ð®ðŽð²ð»ð ð¯ðð¶ð¹ð±ð²ð¿ð, the verdict: ðð²ð²ðœðŠð²ð²ðž ð©ð° ð³ðŒð¿ ð°ðŒðð, ðð£ð§-ð±.ð± ð³ðŒð¿ ð°ð®ðœð®ð¯ð¶ð¹ð¶ðð, ð€ðð²ð»ð¯.ð² ð³ðŒð¿ ð¹ðŒð°ð®ð¹ ð±ð²ðœð¹ðŒð. ðð²ð²ðœðŠð²ð²ðž ð©ð° ð¶ð ð¯ð²ðð ð³ðŒð¿ ðµð¶ðŽðµ-ððŒð¹ððºð², ð°ðŒðð-ðð²ð»ðð¶ðð¶ðð² ð¥ðð. ⢠FLOPs at 27% of V3.2, KV cache at 10% on 1M context (V4-Pro) ⢠V4-Flash pushes further: 10% FLOPs, 7% KV cache ⢠API $1.74/M in, $3.48/M out ðð£ð§-ð±.ð± ð¶ð ð¯ð²ðð ð³ðŒð¿ ð®ðŽð²ð»ðð ððµð®ð ð»ð²ð²ð± ððŒðœ-ðð¶ð²ð¿ ð¿ð²ð®ððŒð»ð¶ð»ðŽ ð®ð»ð± ððŒðŒð¹ ððð². ⢠Terminal-Bench 2.0 at 82.7%, GPQA Diamond at 93.6% ⢠Native multimodal with computer control ⢠API $5/$30 standard, $30/$180 Pro (doubled vs 5.4) ð€ðð²ð»ð¯.ð² ð¶ð ð¯ð²ðð ð³ðŒð¿ ðœð¿ð¶ðð®ðð² ðŒð¿ ðŒð»-ðœð¿ð²ðºð¶ðð² ð±ð²ðœð¹ðŒðð. ⢠35B total / 3B active per token, open source, free to self-host ⢠262K context (extendable to 1M via YaRN) ⢠Native multimodal including video
é®é®è¿ç¯å 容
åçä» åºäºæ¬ç¯ææSkill å
é¢åæš¡æ¿ïŒäžé®äº§åºç»æåç¬è®°æèèµé·èŸŸå
æäžæ¡èèµ / åææ°é»æŽçææèµäººè§è§çé·èŸŸå¡ïŒäº€æèŠç¹ãå€æãç«äºæ Œå±ãé£é©ãå°œè°æž åã
- · 亀æèŠç¹ïŒå ¬åž / 蜮次 / éé¢ / æèµäºº / 䌰åŒïŒæææªæ瀺åå âæªæ«é²âïŒ
- · æèµ thesisïŒè¿å®¶å ¬åžäžºä»ä¹åŒåŸå ³æ³šïŒ
- · ç«äºæ Œå±äžæ¿ä»£æ¹æ¡