News & Updates
December 1, 2025
DeepSeek releases the official version of V3.2: general capabilities are enhanced, and inference-to-Speciale variants are simultaneously launched
On December 1, 2025, DeepSeek released the official version of V3.2 under the MIT protocol, precipitating efficiency exploration such as sparse attention into the official version, and simultaneously launched the DeepSeek-V3.2-Speciale variant that focuses on reasoning, forming a "universal + reasoning" combination.
Source: DeepSeek official
August 21, 2025
DeepSeek releases V3.1: Introducing hybrid reasoning architecture, one-click switching between thinking/non-thinking dual modes
On August 21, 2025, DeepSeek released V3.1 under the MIT protocol. It adopts a thinking/non-thinking dual-mode hybrid architecture and additionally trains over 800B tokens based on V3. It has greatly improved compared to the previous generation on SWE-bench, Terminal-bench and other benchmarks, and enhanced Agent capabilities.
Source: DeepSeek official
September 29, 2025
DeepSeek releases V3.2-Exp experimental version: introducing sparse attention DSA to improve long text efficiency
On September 29, 2025, DeepSeek released the experimental version V3.2-Exp, which introduced DeepSeek Sparse Attention (DSA) to improve computing efficiency in long context scenarios, paving the way for efficiency optimization in subsequent official versions.
Source: DeepSeek official
May 6, 2024
DeepSeek releases V2 open source large model: the first MLA architecture, 236B MoE focusing on high cost performance
In May 2024, DeepSeek released the DeepSeek-V2 series of open source large models, which introduced multi-head latent attention (MLA) and MoE sparse architecture for the first time. It has 236B total parameters and a single token activation of about 21B. It benchmarks mainstream models with extremely low inference costs and lays an efficient route for subsequent V3 and R1.
Source: DeepSeek official
January 20, 2025
DeepSeek releases R1 inference model: MIT open source benchmark OpenAI o1, triggering global AI shock
On January 20, 2025, DeepSeek open sourced the DeepSeek-R1 inference model under the MIT protocol. It was initialized from V3-Base and trained with reinforcement learning-driven training. The mathematical and programming reasoning was benchmarked against OpenAI o1 and is available for free. The application quickly topped the US App Store and caused a shock in the market.
Source: DeepSeek official
April 24, 2026
DeepSeek releases version 3.2.0 to enhance inference stability and API performance
DeepSeek releases 3.2.0, focusing on upgrading complex inference stability and high-concurrency service performance.
Source: DeepSeek official
May 28, 2025
DeepSeek releases upgraded version R1-0528: stronger reasoning, fewer hallucinations, significantly improved mathematical benchmarks
On May 28, 2025, DeepSeek released the upgraded version R1-0528 under the MIT license. The reasoning ability is significantly improved, hallucinations are reduced, and the performance on mathematical benchmarks such as AIME 2025 is significantly enhanced. It is an important iteration of the R1 inference model.
Source: DeepSeek official
May 7, 2026
DeepSeek open source family bucket inventory: turning MoE inference engineering from "black box" to "textbook"
DeepSeek's 2025-2026 intensive open source inference and training infrastructure, FlashMLA, DeepEP, DeepGEMM, DeepSpec, 3FS, etc., covers the entire "algorithm-kernel-communication-storage" link, is open under the MIT protocol, and systematically exports efficient MoE inference engineering capabilities to global developers.
Source: DeepSeek GitHub official organization
June 9, 2026
Understand DeepL Write in one article: AI assistant for multi-language writing and polishing
DeepL Write's core capabilities, cost structure, version evolution, application scenarios and applicable boundaries help the team complete pilots and pre-purchase evaluations.
Source: DeepL Write official documentation
DataRobot capability inventory: Selection reference for AI data processing teams
DataRobot is an enterprise-level AI/ML automation platform founded in 2012. It serves 1,000+ large enterprise customers around the world and provides complete lifecycle solutions from AutoML to model deployment, monitoring alarms and AI governance.
Source: DataRobot official documentation
D-ID publishing side capability inventory: the implementation path of AI video and interactive Avatar
D-ID provides AI video, interactive avatar, video translation and API access capabilities to enterprises and teams. Public information shows that its product line has covered multiple commercial delivery scenarios.
Source: D-ID official documentation, digital human and video generation platform evaluation, AI video tool matching
Understand DALL-E in one article: OpenAI text-to-image model
DALL-E 3 is an OpenAI text-to-image model that can be used directly in ChatGPT. It has accurate prompt word understanding and built-in content security filtering. The API price starts at $0.040/piece, and is suitable for developers to integrate AI image capabilities.
Source: DALL-E official documentation
Cursor Product Observation: Capability Inventory in AI Programming Direction
Cursor is an AI native code editor produced by Anysphere, providing Agent, Composer, Tab, Cmd-K, Bugbot, CLI, Cloud Agents, MCP and other capabilities.
Source: Cursor official documentation
CustomPod.io function dismantling and applicable scenario analysis
CustomPod.io’s public functions, cost structure, version information and applicable boundaries are organized.
Source: CustomPod.io official documentation
CrewAI function dismantling and applicable scenario analysis
CrewAI is a multi-agent orchestration framework that supports role division and collaboration to complete complex tasks. This article summarizes its functions, pricing and application scenarios.
Source: CrewAI official documentation
How to use Craiyon: with executable task flow as the core
Craiyon’s public functions, cost structure, version information and applicable boundaries are organized.
Source: Craiyon official documentation
CowAgent capability inventory: Selection reference for AI agent teams
CowAgent provides terminal agent capabilities and supports multiple models, memory, tool calling and skill expansion. This article sorts out its version rhythm, cost structure, implementation methods and applicable boundaries.
Source: CowAgent official documentation
Coze capability analysis: How to use AI agent scenarios
Coze is a code-free AI Bot construction platform launched by Bytedance. It uses top models such as GPT-4 for free. It has 100+ built-in plug-ins and knowledge base functions, and supports one-click publishing to Feishu, Discord and other channels. The domestic version is "button".
Source: Coze official documentation
Coqui capability analysis: How to use AI audio scenes
Coqui's public functions, cost structure, version information and applicable boundaries are organized.
Source: Coqui official documentation
February 3, 2026
DeepSeek open source OCR-2: Re-understanding long documents with "visual causal flow"
DeepSeek has released a new generation of open source OCR model DeepSeek-OCR-2 on GitHub. It is open sourced under the Apache-2.0 protocol. The technical solution adopts "Visual Causal Flow" and is better than the first generation in long document OCR and complex layout recognition. The first generation project has received approximately 23,700 stars.
Source: DeepSeek GitHub official repository