Ox Alpha | GLM 3.5 Flash Waa Opus Killer Anthropic Iyo OpenAI War Baa U Yaala

Videogan waxaan ka warramayaa model cusub oo shirkadda Shiinaha ee Zhipu AI soo saartay. Maalmahaan waxaa lagu naanaysay OX Alpha, magaciisa rasmiga ahna waa GLM 3.5 Flash. Waa model multi-modal ah, sawir, video iyo qoraal intaba wuu fahmayaa, context-kiisuna waa 1 million token. Waa model yar oo khafiif ah marka la barbardhigo kuwa kale, 320 billion buu yahay halka kuwa kale ay trillions yihiin, sidaas darteed mishiinada Mac mini oo kale ayaad ku run-gareyn kartaa. Toddobaad buu bilaash ahaa, haddase free-gii wuu ka baxay.

Waxa iga soo jiitay in gebi ahaanba lagu run-gareeyo chips AI ah ee Shiinuhu leeyahay, oo aan la isticmaalayn kuwa Maraykanka ama Taiwan. Waxa kale oo la yaab leh in uu la tartamayo DeepSeek V4, GLM-5, Opus 4.8, GPT-5 Terra iyo Gemini Flash, kuwaasoo ah model-lada dhexdhexaadka ah ee shirkadaha waaweyn. Qaybta terminal-ka wuxuu neck and neck la yahay Gemini-ga ugu dambeeyay iyo GPT-5, points aad u kooban ayaana u dhexeeya. TerminalBench-kana way isku dhow yihiin dhammaan, iyadoo Flash-ku uusan ahayn model-ka ugu weyn ee GLM-ka.

Waxaan sidoo kale ku tijaabiyay project yar, design-kiisu wuu nadiif ahaa, dark mode iyo light mode-kuna si fiican bay u shaqeeyeen. Qaybta qiimaha, token-nadu aad bay u raqiis yihiin, cache input-kuna wuu ka sii raqiisan yahay maadaama qoraalladii hore aan dib loo generate-garaynin. Haddii coding aad isticmaalayso, kani wuu kaa hagaagi karaa halka model-lada kale ay qaali ku yihiin, laakiin weli ma aqaano sida uu token-nadiisa u dhaqaaleeyo. Video-ga xiga waxaan tusi doonaa sida Claude Code loogu isticmaalo.

A five-minute Somali-language (Af-Soomaali) AI news video from Dugsiiye on GLM 3.5 Flash, the model the Chinese company Zhipu AI released and which circulated for a week under the name OX Alpha. It is multi-modal, taking images, video and text, carries a one million token context window, and at 320 billion is small enough to run on machines like a Mac mini, unlike the trillion-scale models. The video notes that it runs entirely on Chinese AI chips rather than American or Taiwanese ones, and puts it against DeepSeek V4, GLM-5, Opus 4.8, GPT-5 Terra and Gemini Flash on the terminal benchmarks and TerminalBench, where the gaps come down to a few points. It also shows a small project built with it and looks at how cheap its input, output and cached tokens are. Free on YouTube.