According to @hyperbolic_labs, UC Berkeley Sky Computing Lab researcher Sijun Tan built LLoCO, a technique that processes 128k context while using 30x fewer tokens. source: Hyperbolic Labs on X It delivers 7.62x faster inference in their reported case study. source: Hyperbolic Labs on X The project was powered by Hyperbolic Labs' NVIDIA H100 GPUs. source: Hyperbolic Labs on X

According to @deepseek_ai, DeepSeek released an API update that sets deepseek-chat to non-thinking mode and deepseek-reasoner to thinking mode, a configuration relevant to agent design and latency-sensitive workflows; the announcement was posted on Aug 21, 2025, and is the stated source. According to @deepseek_ai, both model families now support a 128K context window, expanding long-context processing for tasks such as ingesting large data streams; the announcement was posted on Aug 21, 2025, and is the stated source. According to @deepseek_ai, the update adds Anthropic API format support, which enables developers who use Anthropic-style request schemas to interface with DeepSeek models; the announcement was posted on Aug 21, 2025, and is the stated source. According to @deepseek_ai, Strict Function Calling is supported in the Beta API, allowing structured tool invocation that is directly applicable to function-driven trading bots and execution agents; the announcement was posted on Aug 21, 2025, and is the stated source. According to @deepseek_ai, the post also notes more API resources and a smoother experience, signaling ongoing infrastructure improvements that builders can factor into integration plans; the announcement was posted on Aug 21, 2025, and is the stated source.