DeepSeek News

DeepSeek has released R1-0528, featuring improved benchmark performance, enhanced front-end capabilities, reduced hallucinations, and support for JSON output and function calling. The model is available on the chat platform with no API changes, and weights are open-sourced.

DeepSeek has released V3-0324 with major improvements in reasoning, front-end development, and tool-use capabilities. The model is now under the MIT License, and API usage remains unchanged.

DeepSeek released the R1 model on January 20, 2025, with performance on par with OpenAI-o1, fully open-sourced under MIT License, along with distilled small models and API access.

DeepSeek App launched on 2025/01/15, powered by DeepSeek-V3, free with no ads or in-app purchases. Available on App Store, Google Play, and major Android markets. Features include easy login, cross-platform sync, web search, Deep-Think mode, and file upload.

DeepSeek-V3 is the biggest leap forward yet, featuring 671B MoE parameters (37B activated), 60 tokens/second (3x faster than V2), and training on 14.8T high-quality tokens. API pricing is same as V2 until Feb 8, then $0.27/M input (cache miss), $0.07/M (cache hit), $1.10/M output. Fully open-source models and papers.

DeepSeek releases V2.5-1210 as the final version of the V2.5 series, introducing internet search, improved benchmarks in math, coding, writing, and roleplay, and open-sourcing the model on Hugging Face. The team thanks users and hints at next-gen foundation models.

DeepSeek launches R1-Lite-Preview, achieving o1-preview-level performance on AIME & MATH benchmarks, with real-time transparent reasoning and upcoming open-source release.

DeepSeek officially launched DeepSeek-V2.5, merging DeepSeek-V2-0628's general conversational abilities with DeepSeek-Coder-V2-0724's robust code processing. The model shows significant improvements in writing, instruction-following, and safety alignment, and is now available via web, API, and open-source on HuggingFace.

DeepSeek's new disk-based context caching slashes API costs by up to 90% for repeated inputs. Cache hits cost $0.014 per million tokens. The feature works automatically and is especially beneficial for multi-turn conversations, data analysis, and long prompts. First token latency drops from 13s to 500ms for 128K prompts.

DeepSeek API rolls out a major update with support for JSON Output, Function Calling, Chat Prefix Completion (Beta), 8K max_tokens (Beta), and FIM Completion (Beta). These features target deepseek-chat and deepseek-coder models, enhancing developer flexibility and automation.