DeepSeek-V4-Pro Released — 1.6T Parameters, 1M Token Context, MIT License
DeepSeek has released a preview of the "DeepSeek-V4" series on Hugging Face. The Pro version is a 1.6T-parameter MoE model (49B activated), and the Flash version is a 284B-parameter MoE model (13B activated) — both support a **context length of 1 million tokens**.
What stands out is the architectural innovation. A hybrid attention design combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA) keeps **single-token inference FLOPs at 27% and the KV cache at 10%** compared to DeepSeek-V3.2, even at the 1-million-token setting. Improvements also target both training stability and convergence speed, including reinforced residual connections via Manifold-Constrained Hyper-Connections (mHC) and the adoption of the Muon optimizer.
On benchmarks, DeepSeek-V4-Pro-Max (maximum reasoning mode) records 93.5% (Pass@1) on LiveCodeBench and 95.2% on HMMT 2026 Feb, delivering top-tier results in coding and math. The gap with the frontier of closed models is steadily narrowing. It was pretrained on over 32T tokens and is released under the MIT license.
Anthropic Survey: About Half of Claude Users Say They Can Handle 50% or More of Their Work with AI
In a survey Anthropic conducted with roughly 9,700 Claude users, about half reported that they can **handle 50% or more of their work tasks with AI**. Meanwhile, 26% expect AI to cover 60–90% of their work within the next 12 months.
What's interesting is the divide among respondents. Early-career respondents showed the most concern, while heavier users tended to be more optimistic about their own career outlook. This suggests that the more heavily people use AI tools, the more they tend to see AI as an opportunity rather than a threat. That said, it's worth keeping in mind that this survey covered Anthropic's own users.
Apple Vision Pro Lead Joins OpenAI's Hardware Team
Paul Meade, the vice president who oversaw the Vision Pro headset at Apple, is reportedly leaving the company to **join OpenAI's hardware team**. The move is drawing attention as a hire signaling OpenAI's commitment to serious investment in the hardware space. How AI companies' device development efforts will take concrete shape remains to be seen.
OpenAI Restricts GPT-5.6 Rollout Following Government Request
OpenAI has reportedly **restricted the rollout of GPT-5.6 following a request from the government**. According to TechCrunch, OpenAI has stated that "restrictions like this should not become the norm." The news has sparked active debate in the community from the perspective of balancing regulation and innovation, with renewed emphasis on the importance of local LLMs.
NVIDIA Nemotron 3 Ultra — A 550B Model Tuned for Speed
A technical deep-dive (in Japanese) has been published on NVIDIA's Nemotron 3 Ultra (550B), released in early June. Rather than aiming for the top spot in overall benchmark rankings, the model is characterized by a **design that prioritizes inference speed and practicality**. It takes a different approach from high-intelligence models like Kimi K2 and DeepSeek V4, and looks like a noteworthy option for those who prioritize cost-effectiveness in deployment environments.
---