Image created with gemini-3.1-flash-image-preview with claude-opus-4.7. Image prompt: Using the provided reference images, keep the Bell’s Pass Sonoran Desert vista, midday Arizona light, and the exact brown-post ranger-style sign construction and typography, but change the sign header to bold all-caps ‘DEEPSEEK’ with fictional trail entries like ‘→ Open Weights 2.1 mi’, ‘← Reasoning Ridge 0.7 mi’, and ‘→ Mixture Overlook 3.4 mi’, and replace the WP3 medallion with a small enamel whale badge; on the dusty singletrack just past the sign, press subtle whale-shaped tracks into the dirt as if something large quietly passed through, with saguaro, volcanic rock, and the hazy valley behind keeping everything photorealistic.
I have been testing DeepSeek-V4-Pro with the Pi coding agent. I am mindblown by how well it works out of the box. A few notes: I spent a few hours building an LLM wiki with an agent powered entirely by DeepSeek-V4-Pro on @FireworksAI_HQ inference. This is the first time I
https://x.com/omarsar0/status/2050009901234282649
FT: DeepSeek is in talks for its first fundraising round at a valuation of around $45 billion. FT: China’s largest state-backed semiconductor investment fund, which has invested in YMTC and CXMT, is in talks to lead DeepSeek’s fundraising round.
https://x.com/jukan05/status/2051904572038455634
Introducing nanowhale 🐳! A tiny DeepSeek model fully pretrained by an agent. Inspired by @karpathy’s nanochat, we gave ml-intern the task of training a tiny MoE with all the architectural advancements of DeepSeek v4. To test it end-to-end, it trained a 100M-parameter MoE
https://x.com/cmpatino_/status/2051343930373837125
There’s a serious gap in multimodal models – they work with images, but still reason in language, which isn’t that precise for visual stuff. @deepseek_ai just dropped an idea to solve this: let the model literally point to exact locations in the image while it thinks. They call
https://x.com/TheTuringPost/status/2050597658423927134
Welcome to DS4, a specialized inference engine for DeepSeek v4 Flash.
https://t.co/UrUJz5I2R1 This project would have been impossible without the existence of llama.cpp and GGML and the work of @ggerganov and all the other contributors. Thanks!
https://x.com/antirez/status/2052405820235678175
will be very funny if DeepSeek-Vision crushes V4-Pro on ARC-AGI-2 solely because it has some semblance of spatial reasoning
https://x.com/teortaxesTex/status/2049947128189923625
👀 DeepSeek briefly released (then deleted) a vision model tech report — what did it reveal? 👀 Zhihu contributor 刘聪NLP breaks it down: Core idea: 👉 A new multimodal reasoning framework that embeds spatial pointers (boxes & points) directly into the chain-of-thought • The
https://x.com/ZhihuFrontier/status/2050238000433659958
DeepSeek V4–almost on the frontier, a fraction of the price
https://simonwillison.net/2026/Apr/24/deepseek-v4/
DeepSeek v4 pro (max) scores 48.9% on WeirdML, improving on v4 pro (high) at 46.5%, but still well behind Kimi-k2.6 and GLM-5.1 at 56% and 57%, let alone the closed frontier. These runs, like the previous ones, were through Fireworks AI.
https://x.com/htihle/status/2052042076196335658
Leave a Reply