DeepSeek
DeepSeek is a Chinese AI lab, backed by the High-Flyer hedge fund, that shook the industry with open reasoning models trained at a fraction of Western costs. Its R1 release in early 2025 wiped hundreds of billions off US tech stocks and made training efficiency a global priority. All our DeepSeek coverage is collected here.

DeepSeek open-sources DeepInfer: 155 trillion tokens a day and an 80% margin
DeepSeek has open-sourced the V4.1 Flash inference engine. It claims 155 trillion tokens a day from one cluster, a 96% cache hit rate and about $22 million in daily…

A test for AI analysts: the leader scored 68 out of 100, but only 52.5 on forecasting
Researchers in Shanghai froze all available information at the decision date and tested 11 AI systems on 86 investment analyst tasks. Forecasting is where the…

Researchers Remove the Noise from LLM MTP Training with a Single Line of Code
Soft-MTP cleans up the gradient when an LLM learns to predict several tokens at once. On models from 167M to 1.2B parameters, it gives better perplexity and faster…

4.5x Faster Product Launches and 87% Budget Savings: 3 Secure AI Architectures Under Russia's 152-FZ

ENISA and JRC have started testing Chinese open-weight models, three officials confirm

A small finance model learns from its own mistakes: +3 pp without a teacher

Devin now supports ChatGPT sign-in: Fusion with Astra is 39% cheaper

Shadow AI in the Enterprise: Why 90% of Security Officers Break Their Own Bans and How a 3-Zone Traffic Light System Works
According to research from JumpCloud and Microsoft Work Lab, 80% of office workers already use neural networks at work, with 32% hiding it from management, and 90%…

a16z measures AI spending for the first time: only 4.5% of Americans pay, top 1% spends $900
a16z has for the first time measured how much Americans actually spend on AI. Only 4.5% of users are willing to pay for subscriptions to ChatGPT, Gemini, or Claude.

AgentProofBench: CMU researchers test whether web agent verifiers can be reused

Researchers Find 7 "Dead" Models Among 305 in the Ollama Library: They Can't Write Code

Moonshot AI Unveils Kimi K3.1 Pro, the First Pro Model in the Lineup, After the Kibitz Test

Nvidia in talks to buy Reflection AI, a startup valued at $8 billion

AdaPRA Lifts an 80B Qwen by 24.5 Points on Kaggle-Style Leaderboards
Carnegie Mellon researchers gave a data science agent a supervisor with adaptive checklists. With five attempts, an 80B Qwen gained 24.5 percentile points on…

DeepSeek open-sources its DeepInfer inference engine, and analysts calculate an 82% margin
DeepSeek has released its internal DeepInfer framework and the deepseek-recipe component under Apache 2.0. Third-party estimates put the gross margin of its AI…









