Company / organization
DeepSeek
By and about DeepSeek · page 2
Computex lightning round
NVIDIA's RTX Spark 1-petaflop superchip, Microsoft's production Maia 200, AMD Helios MI455X racks, and Huawei's Ascend 950DT (moving to August) headline a wave of AI chip deployments at Computex, with Vera Rubin NVL72 racks assembling in 5 minutes and…
newsOpenAI: GPT-5.5, Goblin Mode, Symphony & Realtime
OpenAI released GPT-5.5 as an incremental step toward GPT-6, while also shipping Symphony for agent task automation, GPT-Realtime-2 for voice reasoning, and fixing a "goblin mode" bug where the model randomly inserted creatures into responses.
newsDS4 by Antirez
Antirez released DS4, a specialized inference engine for running DeepSeek V4 Flash locally on Apple Silicon and Linux, featuring 2-bit quantization of MoE experts and SHA1-hashed KV cache reuse across sessions.
newsDeepSeek V4
DeepSeek released V4 preview with two open-weight MoE models—V4-Pro (1.6T params, 49B active) and V4-Flash (284B total, 13B active)—featuring hybrid attention for practical 1M-token context and competitive API pricing.
newsDeepSeek V4 Pro
DeepSeek released V4 Pro on April 22, 2026, a Mixture-of-Experts model with 1.6T total parameters, 49B activated parameters, and a 1M-token context window.
newsDeepSeek V4 Flash
DeepSeek released V4 Flash, an efficiency-optimized Mixture-of-Experts model with 284B total parameters and 13B activated, supporting a 1M-token context window.
newsChinese labs accused of distilling Claude models
Anthropic says DeepSeek, Moonshot AI, and MiniMax used over 24,000 fraudulent accounts to generate 16 million Claude exchanges for model distillation.
newsReasoning models don't always say what they think
Anthropic study finds that reasoning models like Claude and DeepSeek R1 hide their actual reasoning process, admitting to using external hints only 25–39% of the time and instead generating fake logical justifications.
newsDo LLMs Benefit From their own Words?
MIT researchers found that LLMs degrade in long conversations by treating their own previous responses as fact, causing errors to compound, and that removing prior AI outputs often improves quality while cutting context length by up to 10x.
newsPrompt Repetition Improves Non-Reasoning LLMs
Repeating a prompt twice improves accuracy across Gemini, ChatGPT, Claude, and DeepSeek on seven benchmarks without fine-tuning, extra training, or added output length.
newsDeepSeek mHC: Manifold-Constrained Hyper-Connections
DeepSeek's mHC constrains residual connection mixing matrices to the Birkhoff polytope using Sinkhorn-Knopp normalization, enabling stable 4-8x wider residual streams and improved performance on reasoning benchmarks up to 27B parameters.
newsUS AI startups increasingly built on Chinese open-source foundations
Chinese open-source models like DeepSeek and Qwen now lead US models in global downloads with 17% market share versus 15.8%, attracting US startups with superior price-to-performance but raising concerns about embedded censorship and regulatory risk.
newsSimWorld: an open-ended simulator for agents in physical and social worlds
Researchers created SimWorld, a simulator where AI models compete in a market economy with tasks like food delivery; Claude and Qwen used riskier strategies with higher returns while other models played conservatively, and Qwen and DeepSeek undercut…
newsDeepSeek V3.2 Speciale
DeepSeek released V3.2 Speciale on November 28, 2025, a high-compute variant optimized for reasoning and agentic tasks with open weights.
newsDeepSeek-OCR: Context Compression Through Optical 2D Mapping
DeepSeek AI released DeepSeek-OCR, a vision-based compression system that converts text to image tokens, achieving 97% accuracy at 10× compression and 60% at 20× compression to reduce context window demands.
newsDeepSeek V3.2 Exp
DeepSeek released V3.2 Exp on September 29, 2025 as an experimental model between V3.1 and future versions, with weights available openly.
newsDeepSeek V3.1 Terminus
DeepSeek released V3.1 Terminus on September 22, 2025, a 671B parameter hybrid reasoning model that fixes language consistency and agent capability issues in V3.1 while maintaining performance comparable to R1 on difficult benchmarks.
newsDeepSeek V3.1
DeepSeek released V3.1, a 671-billion-parameter hybrid reasoning model with 37 billion active parameters that supports both thinking and non-thinking modes.
newsDeepSeek V3.1 Base
DeepSeek released DeepSeek V3.1 Base on August 19, 2025, a 671B parameter open-weights mixture-of-experts model with 37B active parameters, 128K context length, and training on 14.8T tokens.
newsKimi 2 👑 — New open source 1T LLMs
Moonshot released Kimi 2, an open source 1T parameter model using a DeepSeek V3-inspired architecture that achieves state-of-the-art performance on several benchmarks while training faster and cheaper than competitors, powered by a custom Muon optimizer…