Deepseek AI Insights: Type-Specific Updates

Deepseek AI Updates by Year and Month

16 Significant Changes from the Last 6 Months

Date Update Type Description View
25-03-2025 DeepSeek V3-0324 Ranks as Top Coding Model on KCores Benchmark Social Media News DeepSeek has unveiled its latest model, V3-0324, rumored to be the foundation for R2, and it’s already making waves! It snagged the second spot among non-thinking coding models on the KCores benchmark, just behind Claude Sonnet 3.5. This benchmark tests models on four coding tasks, and V3-0324 outshined even some big names like R1, o1, Gemini Pro, and Grok.
25-03-2025 DeepSeek-V3-0324 Debuts with Enhanced AI Reasoning and Coding Power Feature DeepSeek has released DeepSeek-V3-0324, an exciting update to its AI model, now available under the MIT License. This version brings big improvements in reasoning, front-end development, and tool-using skills. For simpler tasks, just switch off the DeepThink feature to use the V3 model smoothly. The API stays the same, making it easy for developers to jump in. You can grab the open-source weights on Hugging Face and start exploring this powerful upgrade today!
17-03-2025 DeepSeek Users Seek Longer Chat Limits and Conversation Grouping Options Social Media News DeepSeek AI has impressed users with its capabilities, but its chat length limit forces restarts that disrupt ongoing topics. Frustrated users are asking for an increased chat length—some even willing to pay—or a feature to group conversations for seamless context retention. These enhancements could elevate the user experience for complex tasks and discussions. Share your thoughts on how DeepSeek can improve in the toolkitly discussion section for Deepseek AI.
01-03-2025 DeepSeek Unveils V3/R1 Inference System on Day 6 of #OpenSourceWeek: Boosting AI Efficiency with High CPC Potential AI Tool Benchmarking On Day 6 of #OpenSourceWeek, DeepSeek showcased its cutting-edge DeepSeek-V3/R1 Inference System, optimized for AI performance with cross-node EP-powered batch scaling, computation-communication overlap, and load balancing. Delivering 73.7k input and 14.8k output tokens per second per H800 node, this system achieves a remarkable 545% cost-profit margin. Aimed at advancing AGI goals, this open-source breakthrough promises significant value for AI developers and businesses seeking scalable, cost-effective AI solutions.
28-02-2025 DeepSeek’s 3FS Debuts at #OpenSourceWeek: Fire-Flyer File System Turbocharges AI Workloads with 6.6 TiB/s Read Throughput AI Innovation Update DeepSeek unveiled its Fire-Flyer File System (3FS) on Day 5 of #OpenSourceWeek, a high-performance parallel file system leveraging modern SSDs and RDMA networks to deliver an astounding 6.6 TiB/s aggregate read throughput in a 180-node cluster.
27-02-2025 DualPipe & EPLB Unveiled: DeepSeek Boosts AI Training Efficiency AI Innovation Update DeepSeek’s Day 4 of #OpenSourceWeek introduces DualPipe, a bidirectional pipeline algorithm, and EPLB, an expert-parallel load balancer, for V3/R1 AI training. These tools optimize computation-communication overlap, enhancing GPU efficiency. Explore the open-source AI breakthroughs now!
26-02-2025 DeepSeek API Slashes Prices: 75% Off R1, 50% Off V3 in Off-Peak Hours Offers DeepSeek API introduces off-peak discounts, cutting DeepSeek-R1 costs by 75% and DeepSeek-V3 by 50% daily from 16:30–00:30 UTC. Save big on advanced AI tools for smarter resource use.
26-02-2025 DeepGEMM Launched: High-Speed FP8 AI Library Powers DeepSeek V3/R1 AI Innovation Update DeepGEMM, unveiled on Day 3 of #OpenSourceWeek, is an FP8 GEMM library hitting 1350+ TFLOPS on Hopper GPUs for V3/R1 AI training. With JIT compilation and a lean 300-line core, it outshines expert-tuned tools. Boost your AI projects
25-02-2025 DeepSeek AI Launches DeepEP: Open-Source EP Library for MoE Models on Day 2 of #OpenSourceWeek AI Innovation Update DeepSeek AI unveils DeepEP, the first open-source Expert Parallelism (EP) communication library for Mixture of Experts (MoE) models, boosting training and inference with high-throughput, low-latency features. Available now on GitHub as part of #OpenSourceWeek.
24-02-2025 DeepSeek AI Launches FlashMLA: Open-Source MLA Kernel for Hopper GPUs on Day 1 of #OpenSourceWeek AI Innovation Update DeepSeek AI introduces FlashMLA, an open-source Multi-Latent Attention (MLA) decoding kernel optimized for Hopper GPUs, now in production. With BF16 support, paged KV cache, and blazing performance, it’s available on GitHub as part of #OpenSourceWeek.
21-02-2025 Open-Source Week Kicks Off with DeepSeek AI’s AGI Exploration Company News Starting next week, the tiny team at DeepSeek AI will launch #OpenSourceWeek, sharing 5 battle-tested repositories from their online service. Focused on AGI (Artificial General Intelligence), this open-source AI initiative offers production-ready AI development tools and machine learning repositories. With full transparency, DeepSeek AI’s community-driven effort aims to accelerate AI innovation. Daily unlocks begin soon, blending garage-energy vibes with AI community collaboration. Stay tuned for transparent AI deployment and progress!
18-02-2025 Deepseek Introduces NSA: Ultra-Fast Sparse Attention for Long-Context Training & Inference Company News Deepseek unveils NSA (Natively Trainable Sparse Attention), a breakthrough mechanism designed for ultra-fast long-context training and inference. NSA combines dynamic hierarchical sparse strategies, token compression, and fine-grained selection to optimize performance on modern hardware. It accelerates inference and reduces pre-training costs, delivering top-tier results on general benchmarks, long-context tasks, and instruction-based reasoning.
16-02-2025 DeepSeek Pauses AI App Downloads in South Korea Over Privacy Concerns Company News DeepSeek has halted downloads of its chatbot apps in South Korea following privacy concerns raised by regulators. South Korean authorities are working with the company to enhance data transparency and protection before a potential relaunch.
14-02-2025 Optimize Your DeepSeek-R1 Experience with Official Prompts Service DeepSeek users can now elevate their AI interactions by using the platform’s recommended prompts. These prompts are designed to streamline both searches and file uploads for better performance. Access the official prompts here and make the most out of your DeepSeek-R1 experience!
14-02-2025 Follow DeepSeek Guidelines to Improve AI Accuracy Service Ensure optimal results with DeepSeek-R1 by following the latest guidelines to mitigate model bypass thinking. These tips are tailored to enhance AI responses and improve overall accuracy. Explore the full guide here and take your AI experience to the next level!
20-01-2025 DeepSeek-R1 Launches with Open-Source Excellence Company News DeepSeek-R1, a high-performance, fully open-source AI model, is now live with MIT licensing for free use and commercialization.