DeepSeek 비전 weights 공개, Hy4·로컬 에이전트 성능, Agent 중심 개발과 AI 인프라 투자

Congrats @deepseek_ai on releasing the weights for DeepSeek-V4-Flash-Vision-Exp! HF link: https:// huggingface.co/deepseek-ai/De epSeek-V4-Flash-Vision-Exp … We're working on support now, with an updated cookbook coming ASAP! Welcome to a new era of native vision from DeepSeek

Weights open: https:// huggingface.co/deepseek-ai/De epSeek-V4-Flash-Vision-Exp …

Codex can now run Tencent’s Hy4 preview. - 770B total. - 49B active per token. - 1M context. - Tencent’s blind eval had 163 experts score 203 engineering tasks. - Hy4 came out ahead of GLM 5.3 and Kimi K3. - Public SWE-Bench Pro sits at 65.7. - It is in the Codex picker next to Sol 5.6. - and the $10 plan serves it with no extra API key. https:// x.com/ziwenxu_/statu s/2093425467105722647/video/1 …

A useful number dropped for people running models at home. DeepSeek-V4-Flash EXL3-REAP scored 75.28% on Terminal-Bench 2.1. - in Independent run - 89 tasks. - Losing 7 points and keeping the model on a Spark-class box is the actual news. Highest EXL3’d it to 3bpw, and still hit 75.28% on TB2.1 inside 128GB this score I’ve seen for any model that fits in 128GB. This is the local-agent story, not the leaderboard screenshot story. - http:// github.com/hw3150cu/TB2.1 _penguin_dsv4_flash …

This paper makes you feel that discussions about AI will replace programmers might be heading in completely the wrong direction. The emergence of AI Agents isn't about making software engineers more efficient at their jobs, but about rendering the very act of permanently encoding decision logic into software increasingly unnecessary. What the author is describing is a more fundamental paradigm shift: The essence of traditional software engineering is humans hardcoding judgment logic into code if-else statements, state machines, algorithms these are all ways of pre-baking human decisions. But in Agent systems with LLMs as the core reasoning engine, code becomes a temporarily generated, use-and-discard tool. For each task, the Agent dynamically generates the needed code, executes it, and then discards it. Decisions are no longer pre-encoded but dynamically produced through LLM reasoning loops at runtime. This isn't incremental improvement it's a structural replacement of the software production paradigm. think the most noteworthy detail in this perspective is: This isn't just an upgrade to productivity tools; it's a transformation in the role of software as a concept itself. In the past, code was the center of the system, with Agent frameworks as the shell. Now, the LLM reasoning loop is the center, and code has become a temporary auxiliary inside that shell. If this trend continues, the core competency of software engineering might no longer be writing good code, but designing reliable reasoning constraint boundaries. In the past, focused on writing maintainable code; in the future, might need to focus more on designing reliable reasoning boundaries. Paper - http:// arxiv.org/html/2606.0560 8v1 …

How do frontier labs keep improving a model after the usual post-training stack? One important method is on-policy distillation using the model’s own rollouts plus denser supervision. Here is clear 30-minute visual walkthrough covering: - On-policy vs off-policy - Forward KL vs Reverse KL vs JSD - When to use RL, OPD, SFT or DPO - Privileged information in self-distillation On-policy distillation is becoming one of the more useful techniques in the final stages of LLM training. https:// x.com/neural_avb/sta tus/2092922819390869839/video/1 …

One of the great challenges for British start-ups has been selling to our own Government - until now. Gov procurement has often discriminated against innovative new British companies. @UKSovereignAI has set out to change that, and after a lot of work, we've just launched our rapid procurement scheme - the first of its kind in the UK. This is £100M of contracts for British AI companies to deliver services to make our public services cheaper and more efficient. This scheme is designed around the needs of start-ups: No turnover minimums. Departmental champions from day 1. Upfront payments. Streamlined applications. You keep your IP. We are launching with four initial schemes, with more to come later this year. These are: - Providing AI for a more productive NHS - Integrating AI across Defence environments - Making public compute more efficient - Safe adoption of AI agents. If you think you might bid for one of these, you can first check your eligibility and learn more about the schemes on the website below

We just signed one of the largest AI infrastructure deals for open source, period. 250MW data center. $5B+ in annualized revenue. Built with @HUMAIN in Saudi Arabia.

Open Source algorithm shrunk 31GB of memory down to 4GB TurboVec is a new open-source tool that stores the data your AI app searches through, using 16x less memory. It runs on Google's TurboQuant, which skips the slow setup step every other tool needs. - Faster search than the popular alternative (FAISS) - Works on both Mac and standard servers - Narrow results to exactly what you want - Plugs straight into LangChain and LlamaIndex Your data never leaves your machine. Runs fully offline, works with Python out of the box. - http:// github.com/RyanCodrai/tur bovec …

Your agent says the sweep converged. Pics or it didn't happen. @CoreWeave ARIA now renders live W&B panels in the chat. Line plots and bar charts from your actual run data land right under its analysis, full screenable, with run hover synced to your Workspace.

From image to motion. Created by @GRACEncxl on @MotionMux , The Inheritance begins with Qwen Image 3.0 Pro and comes to life with Wan 3.0—showcasing a complete creative workflow from image to cinematic video.