Web 3.0 Ethiopia - DeFi & AI

Moonshot AI Releases Full Model Weights for the Massive 2.8T Parameter Kimi K3 Moonshot AI has officially released the full model weights and a detailed technical report for Kimi K3, an immense 2.8-trillion parameter Mixture-of-Experts (MoE) model. Activating 104 billion parameters during inference, the model features native vision capabilities and a massive 1-million-token context window. Through cutting-edge architectural innovations such as Kimi Delta Attention, Attention Residuals, and highly optimized MoE routing utilizing 896 specialized experts, Kimi K3 successfully delivers approximately 2.5 times better intelligence per compute unit compared to its predecessor. ሙንሾት ኤአይ (Moonshot AI) 2.8 ትሪሊዮን መለኪያዎች (parameters) ያለውን እጅግ ግዙፍ "ኪሚ ኬ3" (Kimi K3) የተሰኘ ሞዴል ሙሉ የውስጥ አሰራሩን እና ቴክኒካዊ ሪፖርቱን ለህዝብ ክፍት (open-source) አድርጓል። ስራዎችን በሚያከናውንበት ጊዜ (inference) 104 ቢሊዮን ፓራሜትሮችን የሚጠቀመው ይህ ሞዴል፣ ምስልን የመረዳት አቅም እና 1 ሚሊዮን ቶከን መረጃን በአንድ ጊዜ የማንበብ አቅም አለው። ሞዴሉ እንደ "Kimi Delta Attention" እና 896 ልዩ ባለሙያዎችን (experts) በማሳተፍ የተሰራውን የ"MoE" አርክቴክቸር በመጠቀሙ፣ ከቀድሞው ስሪት በ2.5 እጥፍ የተሻለ የአቅም እና የኃይል ቅልጥፍናን አሳይቷል።

Addis AI Launches Comprehensive SDK Addis AI has officially released the Addis AI SDK, an all-in-one software development kit designed to streamline the creation of AI products tailored for African languages. This unified package provides developers with seamless access to speech generation, voice tools, language models, and translation APIs in a single integration. Built specifically for developers focusing on localized AI solutions from day one, the SDK includes support for local languages like Amharic, enabling the rapid development of culturally relevant, voice-driven applications. አዲስ ኤአይ (Addis AI) በተለይ ለአፍሪካ ቋንቋዎች የተሰሩ የሰው ሰራሽ አስተውሎት (AI) ምርቶችን በቀላሉ ለመገንባት የሚያስችለውን አዲስ የሶፍትዌር ማበልጸጊያ (Addis AI SDK) በይፋ አቅርቧል። ይህ የተጠቃለለ መሳሪያ፣ የድምፅ መፍጠሪያዎችን፣ የቋንቋ ሞዴሎችን እና የትርጉም አገልግሎቶችን (APIs) በአንድ መተግበሪያ ውስጥ ያቀርባል። ከጅምሩ የሀገር በቀል ቋንቋዎችን መሰረት አድርገው ለሚሰሩ አልሚዎች (developers) ታስቦ የተነደፈው ይህ ቴክኖሎጂ፣ አማርኛን ጨምሮ የተለያዩ የሀገር ውስጥ ቋንቋዎችን የሚደግፍ ሲሆን፣ በድምፅ የሚታገዙ እና ባህልን የሚያንፀባርቁ መተግበሪያዎችን በፍጥነት ለመስራት ትልቅ እገዛ ያደርጋል። @webthreeth

Anthropic Unveils Claude Opus 5 Delivering Frontier Intelligence at Half the Cost Anthropic has announced the release of Claude Opus 5, a highly proactive and thoughtful AI model that approaches the frontier level intelligence of Fable 5 at approximately half the cost. Excelling in coding, advanced knowledge work, and complex agentic tasks, Opus 5 has established new state of the art records across rigorous industry benchmarks. Most notably, it scored three times higher than the next best model on ARC-AGI-3 while also dominating evaluations like OSWorld 2.0, AutomationBench, and Frontier-Bench. አንትሮፒክ (Anthropic) እጅግ አስተዋይ የሆነውን እና የ"ፌብል 5"ን (Fable 5) አቅም በግማሽ ዋጋ የሚስተካከለውን "ክሎድ ኦፐስ 5" (Claude Opus 5) የተሰኘ አዲስ ሞዴል ይፋ አድርጓል። በኮዲንግ፣ በእውቀት ስራዎች እና በኤጀንት ተግባራት ላይ የላቀ ብቃት ያለው ይህ ሞዴል በተለያዩ ጥብቅ መለኪያዎች ላይ አዲስ የዓለም ክብረ ወሰን አስመዝግቧል። በተለይም በ"ARC-AGI-3" መለኪያ ላይ ከተከታዩ ምርጥ ሞዴል በሶስት እጥፍ የሚበልጥ ውጤት ያመጣ ሲሆን በ"OSWorld 2.0"፣ በ"AutomationBench" እና በ"Frontier-Bench" ፈተናዎች የበላይነቱን ይዟል። @webthreeth

Ethiopian Startup Ambalay Maps Pioneers GeoAI for Landmark-Based Navigation Ambalay Maps, a rising Ethiopian GeoAI startup, is developing location-intelligence tools designed specifically for environments where conventional mapping and formal addressing systems often fall short. By leveraging artificial intelligence, the company offers natural-language navigation in both Amharic and English, utilizing familiar local landmarks instead of relying strictly on street names. Beyond navigation, Ambalay Maps applies AI to analyze satellite imagery for efficient infrastructure identification and offers advanced routing, ETA prediction, and live trip tracking for logistics businesses. በኢትዮጵያ የተመሰረተው "አምባላይ ማፕስ" (Ambalay Maps) የተሰኘ አዲስ የጂኦ-ኤአይ (GeoAI) ስታርትአፕ፣ መደበኛ የካርታ እና የአድራሻ ስርዓቶች በሚያዳግቱባቸው ቦታዎች የሚያገለግል የካርታ ቴክኖሎጂ እያበለጸገ ይገኛል። የሰው ሰራሽ አስተውሎትን (AI) በመጠቀም፣ ድርጅቱ በመንገድ ስሞች ብቻ ከመመራት ይልቅ ሰዎች በሚያውቋቸው ታዋቂ ቦታዎች (landmarks) ላይ በመመስረት በአማርኛ እና በእንግሊዝኛ ቋንቋዎች አቅጣጫን የሚጠቁም ቴክኖሎጂ አቅርቧል። ከአቅጣጫ ጠቋሚነት በተጨማሪ፣ ህንፃዎችን እና መንገዶችን በፍጥነት ለመለየት የሳተላይት ምስሎችን በኤአይ የሚተነትን ሲሆን፣ ለሎጂስቲክስ እና የዕቃ ማጓጓዣ (delivery) ተቋማት የተሻለ የመንገድ ምርጫን፣ የጊዜ ትንበያን እና የቀጥታ ጉዞ ክትትልን ያቀርባል። @webthreeth

OpenAI has shared new growth data highlighting a massive surge in the adoption of its autonomous platforms, describing the trend as an "agentic breakout." According to a recently published chart, the combined weekly active users for ChatGPT Work and Codex skyrocketed from 6 million on July 12 to 10 million by July 21, 2026. This explosive growth represents nearly a two-fold increase in just over a single week, underscoring the rapidly accelerating enterprise demand for AI-driven workflows and autonomous coding assistants following their recent launches. ኦፕንኤአይ (OpenAI) "የኤጀንት ቴክኖሎጂ ፈጣን እድገት" (agentic breakout) ሲል የገለጸውን እና የኤአይ መድረኮቹ ተጠቃሚዎች በከፍተኛ ፍጥነት መጨመራቸውን የሚያሳይ አዲስ መረጃ ይፋ አድርጓል። ይፋ በተደረገው መረጃ መሰረት፣ የ"ቻት ጂፒቲ ዎርክ" (ChatGPT Work) እና "ኮዴክስ" (Codex) ሳምንታዊ ተጠቃሚዎች ቁጥር በጁላይ 12 ከነበረበት 6 ሚሊዮን፣ በአንድ ሳምንት ጊዜ ውስጥ ብቻ በጁላይ 21 ወደ 10 ሚሊዮን ከፍ ብሏል። በአንድ ሳምንት ውስጥ ወደ እጥፍ የተጠጋ ጭማሪ ያሳየው ይህ አሃዝ፣ በዓለም ዙሪያ ለኤአይ የቢሮ ስራ ወኪሎች እና የኮዲንግ ረዳቶች ያለውን እጅግ ከፍተኛ እና ፈጣን ፍላጎት በግልጽ የሚያሳይ ነው። @webthreeth

Ethiopian-Built ScholarXIV Tops Global Agent-Readiness Leaderboard with Perfect Score ScholarXIV, an Ethiopian-built AI platform designed for searching, analyzing, and interacting with arXiv research papers, has achieved a historic milestone by scoring a perfect 100/100 on Farming Labs' agent-readiness benchmark. This flawless score places ScholarXIV's developer documentation at the #1 spot on the global public leaderboard, successfully outperforming major international platforms such as Mintlify (94), Notion (93), and Anthropic's Claude (92). የሳይንሳዊ ጥናት ወረቀቶችን (arXiv research papers) ለመፈለግ፣ ለመተንተን እና ከእነሱ ጋር በኤአይ (AI) ለመወያየት የሚያስችለውና በኢትዮጵያውያን የተገነባው ScholarXIV የተሰኘ መድረክ፣ በ'Farming Labs' የኤአይ ወኪል ዝግጁነት መለኪያ (agent-readiness benchmark) ሙሉ 100/100 በማምጣት ትልቅ የታሪክ ምዕራፍ አስመዝግቧል። በዚህ ፍጹም ውጤት፣ የScholarXIV የገንቢዎች ሰነድ (developer documentation) በዓለም አቀፍ ደረጃ አንደኛ በመውጣት እንደ ሚንትሊፋይ (Mintlify - 94)፣ ኖሽን (Notion - 93) እና የአንትሮፒኩ ክሎድ (Claude - 92) ያሉ ግዙፍ ዓለም አቀፍ ተቋማትን አሸንፏል። @webthreeth

Google AI has officially announced Gemini 3.6 Flash, a new model that delivers higher intelligence, improved real-world usability, and lower pricing based directly on developer feedback. A major highlight of this release is its remarkable token efficiency. Recent efficiency data reveals a sharp reduction in the number of output tokens required to complete individual tasks, making the model significantly faster and more cost-effective for developers building scalable AI applications. የጎግል ኤአይ (Google AI) በአልሚዎች አስተያየት ላይ በመመስረት የላቀ ብልህነትን፣ ዝቅተኛ ዋጋን እና የተሻለ የተግባር አጠቃቀምን ያካተተውን "Gemini 3.6 Flash" (ጄሚኒ 3.6 ፍላሽ) የተሰኘ አዲስ ሞዴል ይፋ አድርጓል። የዚህ ማሻሻያ ዋና መገለጫ የቶከን (token) አጠቃቀም ቅልጥፍናው ሲሆን፣ መረጃዎች እንደሚያሳዩት ሞዴሉ ለእያንዳንዱ ተግባር የሚያወጣው የውጤት ቶከን መጠን በከፍተኛ ሁኔታ ቀንሷል። ይህም አዲሱን ሞዴል ይበልጥ ፈጣን እና ወጪ ቆጣቢ በማድረግ፣ የሰው ሰራሽ አስተውሎት መተግበሪያዎችን ለሚገነቡ ባለሙያዎች እጅግ ተመራጭ ያደርገዋል። @webthreeth

Google Unveils Gemini 3.5 Flash-Lite: Its Smallest and Fastest Model Yet Google AI has introduced Gemini 3.5 Flash-Lite, officially its smallest and fastest model to date. Designed to replace the outgoing 2.5 Flash model at the exact same price point, this new iteration delivers higher overall intelligence than Gemini 3 and outperforms the 3.1 Flash-Lite on most tasks. Capable of generating nearly 350 output tokens per second, the model is perfectly optimized for latency-sensitive user interfaces and provides a highly efficient, low-cost option for driving complex AI agent frameworks. የጎግል ኤአይ (Google AI) እስካሁን ካመረታቸው ሞዴሎች እጅግ አነስተኛ እና ፈጣን የሆነውን "Gemini 3.5 Flash-Lite" (ጄሚኒ 3.5 ፍላሽ-ላይት) በይፋ አስተዋውቋል። ቀስ በቀስ ከገበያ እየወጣ ባለው የ"2.5 Flash" የዋጋ መጠን የሚቀርበው ይህ አዲስ ሞዴል፣ ከ"ጄሚኒ 3" የላቀ ብልህነት ያለው ሲሆን በአብዛኛዎቹ ተግባራት ላይ ከ"3.1 Flash-Lite" የተሻለ አቅም አሳይቷል። በሰከንድ ወደ 350 ቶከኖችን (tokens) የማመንጨት ከፍተኛ ፍጥነት ያለው በመሆኑ፣ ፈጣን ምላሽ ለሚፈልጉ መተግበሪያዎች (latency-sensitive interfaces) እና የኤአይ ወኪሎችን (AI agents) ስርዓት በብቃት ለማስተዳደር እጅግ ተመራጭ እና ወጪ ቆጣቢ አማራጭ ነው። @webthreeth

Alibaba Unveils Qwen3.8: A Massive 2.4 Trillion Parameter Multimodal Model Alibaba's Qwen team has announced the upcoming release of Qwen3.8, a groundbreaking 2.4 trillion parameter multimodal AI model. Set to be launched as an open-weight model, this release continues the growing trend of Chinese AI laboratories delivering exceptionally large-scale, open-source AI infrastructure to the global community. የአሊባባ (Alibaba) የ"Qwen" ቡድን 2.4 ትሪሊዮን ፓራሜትሮች ያሉት "Qwen3.8" (ኩዌን 3.8) የተሰኘ እጅግ ግዙፍ መልቲ-ሞዳል (multimodal) የሰው ሰራሽ አስተውሎት ሞዴል በቅርቡ እንደሚለቅ አስታውቋል። ይህ ሞዴል ለህዝብ ክፍት (open-weight) ሆኖ የሚለቀቅ ሲሆን፣ የቻይና የኤአይ ላቦራቶሪዎች እጅግ ግዙፍ ሞዴሎችን በነጻ የማቅረብ አዝማሚያን አጠናክሮ የቀጠለ ነው። @webthreeth

Benchmark Scores of KIMI K3 Kimi K3 demonstrates strong performance across coding benchmarks, leading in Terminal Bench 2.1 with 88.3, Program Bench at 77.8, and SWE Marathon at 42.0, while securing second place in FrontierSWE (81.2) and Kimi Code Bench (72.9) behind models like Fable 5. In general agent tasks, it excels by topping Automation Bench (30.8), SpreadsheetBench 2 (34.8), and BrowseComp (91.2), with competitive second-place scores in GDPval-AA v2 (1668) and JobBench (52.9). @webthreeth

Moonshot AI’s Kimi-K3 Surges to the Top of the Frontend Code Arena Leaderboard Arena.ai has reported that Moonshot AI’s Kimi-K3 model has secured the top position on the Frontend Code Arena leaderboard with an impressive score of 1,679 points. This remarkable leap saw the model climb from 18th place to surpass Anthropic’s Claude Fable 5, which scored 1,631. Demonstrating its unmatched capability in web development, Kimi-K3 achieved first place in 6 out of 7 frontend evaluation domains, including Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools. "አሬና.ኤአይ" (Arena.ai) ባወጣው አዲስ መረጃ መሰረት፣ የሙንሾት ኤአይ (Moonshot AI) "Kimi-K3" (ኪሚ-ኬ3) ሞዴል 1,679 ነጥብ በማስመዝገብ በድረ-ገጽ የኮዲንግ ውድድር (Frontend Code Arena) አንደኛ ደረጃን ይዟል። ሞዴሉ ከ18ኛ ደረጃ በከፍተኛ ፍጥነት በመመንጠቅ፣ 1,631 ነጥብ ያለውን የአንትሮፒክ "ክሎድ ፌብል 5"ን በማሸነፍ ነው ወደ መሪነት የመጣው። ይህ ቴክኖሎጂ በድረ-ገጽ ልማት ያለውን የላቀ አቅም በማሳየት፣ ከ7ቱ የፍተሻ ዘርፎች በ6ቱ (Brand & Marketing፣ Data & Analytics እና Simulations ጨምሮ) አንደኛ ወጥቷል። @webthreeth

It is really impressive from what I tried. I can only speak about my subject, but it did better than GPT 5.6

Help a brother out and let me try it in depth

Boost me and we both win! Sign up on Kimi and we each get a guaranteed benefit — up to 1-Year Membership Credits: https://kimi-bot.com/activities/viral-referral/share?scenario=invite&from=share_poster&invitation_code=WW7B6J

Kimi K3 has been released on their platform and is available for evaluation. It is considered superior to GPT 5.5 and is confirmed as a frontier model. Users are encouraged to access it via their official website. News coverage will follow upon the official release. @webthreeth

KIMI K3 - RUMORS (FABLE-5 Level Open Source) @webthreeth