DeepSeek Updates and Changelog

DeepSeek released V4 Preview in two variants: V4-Pro for maximum capability and V4-Flash for speed-optimized use cases. Both are open-sourced and available on chat, app, and API. The release included significant improvements in knowledge, reasoning, and agentic capabilities, meaning V4 can perform complex multi-step tasks and workflows more autonomously than previous versions. The Pro and Flash split mirrors the pattern established by other labs (OpenAI's GPT-4o and GPT-4o mini, Anthropic's Opus and Haiku), but DeepSeek offers both tiers at dramatically lower API pricing. V4 arrived months later than analysts expected, reinforcing the pattern of CEO Liang Wenfeng prioritizing quality over speed.

The deepseek-chat and deepseek-reasoner API model names will be discontinued on July 24, 2026. During the transition period, these names point to V4 Flash's non-thinking and thinking modes respectively. Developers should migrate to explicit V4 model IDs.

A significant update to the R1 reasoning model. Benchmark gains across AIME 2025, GPQA, LCB_v6, and Aider. The release specifically targeted hallucination reduction, which had been one of R1's most criticized weaknesses, and improved JSON output and function calling for developer use cases. Complex reasoning tasks consume more tokens compared to the legacy R1 version, a trade-off between accuracy and cost that most users accepted.

V3-0324 delivered sharper reasoning and smarter code generation alongside upgraded Chinese writing and improved translation. Benchmark gains across MMLU-Pro, GPQA, AIME, and LiveCodeBench. The release also brought UI polish and better function calling. This was the version that made V3 genuinely competitive with GPT-4 class models for everyday developer use cases.

The reasoning model that changed the AI industry's cost assumptions. R1 achieved performance comparable to OpenAI's o1 on certain reasoning benchmarks while being trained at a fraction of the reported cost. Released under an MIT license with full model weights. Within a week, the DeepSeek app surpassed ChatGPT as the most downloaded free app on the US iOS App Store. Nvidia lost hundreds of billions in market capitalization as investors reconsidered whether frontier AI performance truly required frontier-scale compute budgets. The moment forced every major AI lab to explain its spending to investors in new ways.

The base model that preceded R1. V3 used a Mixture of Experts architecture to deliver strong general-purpose performance at low inference costs. Released open-source, it became the foundation for the R1 reasoning model that launched a month later. V3 demonstrated that architectural efficiency could close the performance gap with models trained on far larger compute budgets.

DeepSeek launched V3.2-Speciale as a temporary API endpoint with the same pricing as the standard model. The experimental release let developers test specialized capabilities before they were rolled into the main V3 line.

V2.5 combined the general chat and coding capabilities that had previously been split into separate models. Developers no longer needed to route requests to different endpoints based on task type. The unified model maintained strong performance on both conversational and code generation benchmarks.

V2 introduced the Mixture of Experts architecture that would define DeepSeek's approach going forward. The architecture activates only a subset of the model's parameters for each input, dramatically reducing inference costs while maintaining strong benchmark performance. API pricing was set far below competitors, establishing DeepSeek's cost leadership positioning.

DeepSeek's first dedicated code generation model, available in multiple sizes from 1.3B to 33B parameters. Released open-source, it provided a capable alternative to Codex and early Copilot models for developers who needed to self-host.