@sky.wav: Their new V4 Flash 0731 scored 50 on the Intelligence Index (+10 points) — matching GPT-5.6 Luna and beating GLM 5.2, which was the previous SOTA just a month ago. But the insane part? It’s ~70% smaller than GLM 5.2… while outperforming it. Same ultra-efficient architecture: 284B total parameters but only 13B active, full 1 million token context. This means high-end intelligence on a single RTX 5090 is coming way sooner than the 18-month predictions. API pricing is still only $0.14 / $0.28 per million tokens — roughly 60% cheaper than OpenAI even after their big price cut, with 98% cache hit discount making repeated work almost free. Agentic performance? Elo rating jumped from 1189 → 1559. Frontend Coding Arena? 7st overall with 1586 points (+154 from preview), 3st among open models, and the best performance-per-dollar of any model. Token usage down. Hallucinations down. Meanwhile Claude Sonnet is looking expensive and mid Chinese models just delivered better intelligence + much smaller size + lower cost all at once. Local AI just got massively accelerated 🇨🇳 Western labs… are y’all okay?? 👀 #deepseekv4 #chatgpt #aimodel #bestai #aifyp

kissdpups52 :

Lanoz :

DeepSeek just Claude with diff name

Получается Некач :

google seen this

tik_tok_shau :

great in benchmarks, poor in everyday use

Yêu cầu tin nhắn :

best open source ai

￴ ￴ ￴ ￴ ￴ ￴ ￴ ￴ ￴ ￴ ￴ ￴ :

Which Gemini 3.1pro or Gemini 3.6 flash is good for coding?

kexol :

what about 5.6 luna? cheaper and smarter than deepseek

Douche milk 🥛 :

тесты тестами а в итоге модель делает не оптимизированный и не логичный код👍

￴ ￴ ￴ ￴ ￴ ￴ ￴ ￴ ￴ ￴ ￴ ￴ :

y'all are overhyping ts, it isn't even multimodal, it cant see images

_skyrol :

First image is straight up a lie 😭

Andy Wang🇨🇳🇺🇸🇳🇿 :

Just a side product from a quantitative trading company btw

kotyara :

дипсик

RFD :

tidak melihat Gemini 3.1 pro

Rexi :

Ko biết gehihi đang ở ngó nào

LASSDS :

gemimi 3.6 hight yếu ác. tác vụ tìm kiếm phân tích nhanh thôi. chứ để nó làm task dài 1 tid nó làm biếng ác. còn ngoo và hay quên

nokls :

ну и где ваш дипсик

losissiski :

тесты тестами. а повседневное задачи это другое) общение с ним это если головы

PoV :

hella cheap too

tataltaufavorit :

Deepseeek being trained by ChatGPT and Claude models.

Oscar - ****** :

gemini 3.1 pro?

Mangora_Lore :

Он уже доступен api deepseek? Или только у других провайдеров?

adevaratulspigotar :

it's bad

eyuuut :

Там же в4 про есть вроде

. :

cant even see images and focus on the request at same time btw🥀