惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

Apple Machine Learning Research
Apple Machine Learning Research
AWS News Blog
AWS News Blog
Google DeepMind News
Google DeepMind News
U
Unit 42
博客园 - 叶小钗
博客园 - 聂微东
GbyAI
GbyAI
Stack Overflow Blog
Stack Overflow Blog
有赞技术团队
有赞技术团队
aimingoo的专栏
aimingoo的专栏
D
DataBreaches.Net
CTFtime.org: upcoming CTF events
CTFtime.org: upcoming CTF events
Jina AI
Jina AI
美团技术团队
The Cloudflare Blog
M
MIT News - Artificial intelligence
Microsoft Azure Blog
Microsoft Azure Blog
I
InfoQ
S
Schneier on Security
C
Check Point Blog
Project Zero
Project Zero
The Hacker News
The Hacker News
Scott Helme
Scott Helme
K
KPMG report finds enterprise disconnect between AI and its ROI | CIO
Cisco Talos Blog
Cisco Talos Blog
P
Privacy International News Feed
SecWiki News
SecWiki News
Latest news
Latest news
MongoDB | Blog
MongoDB | Blog
S
Secure Thoughts
Google Online Security Blog
Google Online Security Blog
F
Fortinet All Blogs
博客园 - 三生石上(FineUI控件)
H
Help Net Security
TaoSecurity Blog
TaoSecurity Blog
cs.CL updates on arXiv.org
cs.CL updates on arXiv.org
Last Week in AI
Last Week in AI
P
Privacy & Cybersecurity Law Blog
Forbes - Security
Forbes - Security
G
GRAHAM CLULEY
N
Netflix TechBlog - Medium
L
Lohrmann on Cybersecurity
A
About on SuperTechFans
T
The Exploit Database - CXSecurity.com
C
Cisco Blogs
PCI Perspectives
PCI Perspectives
大猫的无限游戏
大猫的无限游戏
T
Troy Hunt's Blog
H
Hacker News: Front Page
Vercel News
Vercel News

MarkTechPost

A Coding Implementation of End-to-End Brain Decoding from MEG Signals Using NeuralSet and Deep Learning for Predicting Linguistic Features Meta Introduces Autodata: An Agentic Framework That Turns AI Models into Autonomous Data Scientists for High-Quality Training Data Creation A Coding Guide on LLM Post Training with TRL from Supervised Fine Tuning to DPO and GRPO Reasoning Qwen AI Releases Qwen-Scope: An Open-Source Sparse AutoEncoders (SAE) Suite That Turns LLM Internal Features into Practical Development Tools A Coding Deep Dive into Agentic UI, Generative UI, State Synchronization, and Interrupt-Driven Approval Flows Moonshot AI Open-Sources FlashKDA: CUTLASS Kernels for Kimi Delta Attention with Variable-Length Batching and H20 Benchmarks Microsoft Research’s World-R1 Uses Flow-GRPO and 3D-Aware Rewards to Inject Geometric Consistency Into Wan 2.1 Without Architectural Changes A Coding Implementation on Pyright Type Checking Covering Generics, Protocols, Strict Mode, Type Narrowing, and Modern Python Typing IBM Releases Two Granite Speech 4.1 2B Models: Autoregressive ASR with Translation and Non-Autoregressive Editing for Fast Inference Top 10 KV Cache Compression Techniques for LLM Inference: Reducing Memory Overhead Across Eviction, Quantization, and Low-Rank Methods Qwen Team Releases FlashQLA: a High-Performance Linear Attention Kernel Library That Achieves Up to 3× Speedup on NVIDIA Hopper GPUs Step by Step Guide to Build a Complete PII Detection and Redaction Pipeline with OpenAI Privacy Filter Meta FAIR Releases NeuralSet: A Python Package for Neuro-AI That Supports fMRI, M/EEG, Spikes, and HuggingFace Embeddings smol-audio: A Colab-Friendly Notebook Collection for Fine-Tuning Whisper, Parakeet, Voxtral, Granite Speech, and Audio Flamingo 3 A Coding Implementation on Document Parsing Benchmarking with LlamaIndex ParseBench Using Python, Hugging Face, and Evaluation Metrics Poolside AI Introduces Laguna XS.2 and M.1: Agentic Coding Models Reaching 68.2% and 72.5% on SWE-bench Verified How to Build Traceable and Evaluated LLM Workflows Using Promptflow, Prompty, and OpenAI OpenAI Releases Privacy Filter: A 1.5B-Parameter Open-Source PII Redaction Model with 50M Active Parameters Top 10 Physical AI Models Powering Real-World Robots in 2026 How to Build a Lightweight Vision-Language-Action-Inspired Embodied Agent with Latent World Modeling and Model Predictive Control Meet Talkie-1930: A 13B Open-Weight LLM Trained on Pre-1931 English Text for Historical Reasoning and Generalization Research Build a Reinforcement Learning Powered Agent that Learns to Retrieve Relevant Long-Term Memories for Accurate LLM Question Answering OpenMOSS Releases MOSS-Audio: An Open-Source Foundation Model for Speech, Sound, Music, and Time-Aware Audio Reasoning Meta AI Releases Sapiens2: A High-Resolution Human-Centric Vision Model for Pose, Segmentation, Normals, Pointmap, and Albedo The LoRA Assumption That Breaks in Production How to Build a Fully Searchable AI Knowledge Base with OpenKB, OpenRouter, and Llama How to Build Smarter Multilingual Text Wrapping with BudouX Through Parsing, HTML Rendering, Model Introspection, and Toy Training Top 7 Benchmarks That Actually Matter for Agentic Reasoning in Large Language Models RAG Without Vectors: How PageIndex Retrieves by Reasoning A Coding Tutorial on Datashader on Rendering Massive Datasets with High-Performance Python Visual Analytics xAI Launches grok-voice-think-fast-1.0: Topping τ-voice Bench at 67.3%, Outperforming Gemini, GPT Realtime, and More A Coding Implementation on kvcached for Elastic KV Cache Memory, Bursty LLM Serving, and Multi-Model GPU Sharing Google DeepMind Introduces Vision Banana: An Instruction-Tuned Image Generator That Beats SAM 3 on Segmentation and Depth Anything V3 on Metric Depth Estimation Meet GitNexus: An Open-Source MCP-Native Knowledge Graph Engine That Gives Claude Code and Cursor Full Codebase Structural Awareness A Coding Implementation on Deepgram Python SDK for Transcription, Text-to-Speech, Async Audio Processing, and Text Intelligence A Coding Implementation on Microsoft’s OpenMementos with Trace Structure Analysis, Context Compression, and Fine-Tuning Data Preparation Google DeepMind Introduces Decoupled DiLoCo: An Asynchronous Training Architecture Achieving 88% Goodput Under High Hardware Failure Rates Mend Releases AI Security Governance Framework: Covering Asset Inventory, Risk Tiering, AI Supply Chain Security, and Maturity Model Mend.io Releases AI Security Governance Framework Covering Asset Inventory, Risk Tiering, AI Supply Chain Security, and Maturity Model OpenAI Releases GPT-5.5, a Fully Retrained Agentic Model That Scores 82.7% on Terminal-Bench 2.0 and 84.9% on GDPval A Coding Tutorial on OpenMythos on Recurrent-Depth Transformers with Depth Extrapolation, Adaptive Computation, and Mixture-of-Experts Routing Google Cloud AI Research Introduces ReasoningBank: A Memory Framework that Distills Reasoning Strategies from Agent Successes and Failures Xiaomi Releases MiMo-V2.5-Pro and MiMo-V2.5: Matching Frontier Model Benchmarks at Significantly Lower Token Cost How to Design a Production-Grade CAMEL Multi-Agent System with Planning, Tool Use, Self-Consistency, and Critique-Driven Refinement Alibaba Qwen Team Releases Qwen3.6-27B: A Dense Open-Weight Model Outperforming 397B MoE on Agentic Coding Benchmarks A Detailed Implementation on Equinox with JAX Native Modules, Filtered Transforms, Stateful Layers, and End-to-End Training Workflows Next Leap to Harness Engineering: JiuwenClaw Pioneers ‘Coordination Engineering’ Photon Releases Spectrum: An Open-Source TypeScript Framework that Deploys AI Agents Directly to iMessage, WhatsApp, and Telegram OpenAI Open-Sources Euphony: A Browser-Based Visualization Tool for Harmony Chat Data and Codex Session Logs Hugging Face Releases ml-intern: An Open-Source AI Agent that Automates the LLM Post-Training Workflow A Coding Implementation to Build a Conditional Bayesian Hyperparameter Optimization Pipeline with Hyperopt, TPE, and Early Stopping Google Introduces Simula: A Reasoning-First Framework for Generating Controllable, Scalable Synthetic Datasets Across Specialized AI Domains A Coding Implementation on Qwen 3.6-35B-A3B Covering Multimodal Inference, Thinking Control, Tool Calling, MoE Routing, RAG, and Session Persistence Moonshot AI Releases Kimi K2.6 with Long-Horizon Coding, Agent Swarm Scaling to 300 Sub-Agents and 4,000 Coordinated Steps A Coding Implementation on Microsoft’s Phi-4-Mini for Quantized Inference Reasoning Tool Use RAG and LoRA Fine-Tuning OpenAI Scales Trusted Access for Cyber Defense With GPT-5.4-Cyber: a Fine-Tuned Model Built for Verified Security Defenders Moonshot AI and Tsinghua Researchers Propose PrfaaS: A Cross-Datacenter KVCache Architecture that Rethinks How LLMs are Served at Scale Meet OpenMythos: An Open-Source PyTorch Reconstruction of Claude Mythos Where 770M Parameters Match a 1.3B Transformer How TabPFN Leverages In-Context Learning to Achieve Superior Accuracy on Tabular Datasets Compared to Random Forest and CatBoost A Coding Implementation to Build an AI-Powered File Type Detection and Security Analysis Pipeline with Magika and OpenAI NVIDIA Releases Ising: the First Open Quantum AI Model Family for Hybrid Quantum-Classical Systems xAI Launches Standalone Grok Speech-to-Text and Text-to-Speech APIs, Targeting Enterprise Voice Developers A Coding Tutorial for Running PrismML Bonsai 1-Bit LLM on CUDA with GGUF, Benchmarking, Chat, JSON, and RAG A Coding Guide for Property-Based Testing Using Hypothesis with Stateful, Differential, and Metamorphic Test Design Anthropic Releases Claude Opus 4.7: A Major Upgrade for Agentic Coding, High-Resolution Vision, and Long-Horizon Autonomous Tasks Google AI Releases Auto-Diagnose: An Large Language Model LLM-Based System to Diagnose Integration Test Failures at Scale A End-to-End Coding Guide to Running OpenAI GPT-OSS Open-Weight Models with Advanced Inference Workflows Top 19 AI Red Teaming Tools (2026): Secure Your ML Models A Coding Guide to Build a Production-Grade Background Task Processing System Using Huey with SQLite, Scheduling, Retries, Pipelines, and Concurrency Control Qwen Team Open-Sources Qwen3.6-35B-A3B: A Sparse MoE Vision-Language Model with 3B Active Parameters and Agentic Coding Capabilities OpenAI Launches GPT-Rosalind: Its First Life Sciences AI Model Built to Accelerate Drug Discovery and Genomics Research Building Transformer-Based NQS for Frustrated Spin Systems with NetKet UCSD and Together AI Research Introduces Parcae: A Stable Architecture for Looped Language Models That Achieves the Quality of a Transformer Twice the Size How to Build a Universal Long-Term Memory Layer for AI Agents Using Mem0 and OpenAI A Coding Implementation to Build Multi-Agent AI Systems with SmolAgents Using Code Execution, Tool Calling, and Dynamic Orchestration A Technical Deep Dive into the Essential Stages of Modern Large Language Model Training, Alignment, and Deployment Google AI Launches Gemini 3.1 Flash TTS: A New Benchmark in Expressive and Controllable AI Voice Google DeepMind Releases Gemini Robotics-ER 1.6: Bringing Enhanced Embodied Reasoning and Instrument Reading to Physical AI Google Launches ‘Skills’ in Chrome: Turning Reusable AI Prompts into One-Click Browser Workflows A Coding Implementation of Crawl4AI for Web Crawling, Markdown Generation, JavaScript Execution, and LLM-Based Structured Extraction TinyFish AI Releases Full Web Infrastructure Platform for AI Agents: Search, Fetch, Browser, and Agent Under One API Key NVIDIA and the University of Maryland Researchers Released Audio Flamingo Next (AF-Next): A Super Powerful and Open Large Audio-Language Model A Hands-On Coding Tutorial for Microsoft VibeVoice Covering Speaker-Aware ASR, Real-Time TTS, and Speech-to-Speech Pipelines Meta AI and KAUST Researchers Propose Neural Computers That Fold Computation, Memory, and I/O Into One Learned Model A Coding Implementation of MolmoAct for Depth-Aware Spatial Reasoning, Visual Trajectory Tracing, and Robotic Action Prediction MiniMax Just Open Sourced MiniMax M2.7: A Self-Evolving Agent Model that Scores 56.22% on SWE-Pro and 57.0% on Terminal Bench 2 Liquid AI Releases LFM2.5-VL-450M: a 450M-Parameter Vision-Language Model with Bounding Box Prediction, Multilingual Support, and Sub-250ms Edge Inference Researchers from MIT, NVIDIA, and Zhejiang University Propose TriAttention: A KV Cache Compression Method That Matches Full Attention at 2.5× Higher Throughput How to Build a Secure Local-First Agent Runtime with OpenClaw Gateway, Skills, and Controlled Tool Execution How Knowledge Distillation Compresses Ensemble Intelligence into a Single Deployable AI Model Alibaba’s Tongyi Lab Releases VimRAG: a Multimodal RAG Framework that Uses a Memory Graph to Navigate Massive Visual Contexts A Coding Guide to Markerless 3D Human Kinematics with Pose2Sim, RTMPose, and OpenSim NVIDIA Releases AITune: An Open-Source Inference Toolkit That Automatically Finds the Fastest Inference Backend for Any PyTorch Model Five AI Compute Architectures Every Engineer Should Know: CPUs, GPUs, TPUs, NPUs, and LPUs Compared An End-to-End Coding Guide to NVIDIA KVPress for Long-Context LLM Inference, KV Cache Compression, and Memory-Efficient Generation Meta Superintelligence Lab Releases Muse Spark: A Multimodal Reasoning Model With Thought Compression and Parallel Agents Sigmoid vs ReLU Activation Functions: The Inference Cost of Losing Geometric Context A Coding Guide to Build Advanced Document Intelligence Pipelines with Google LangExtract, OpenAI Models, Structured Extraction, and Interactive Visualization Google AI Research Introduces PaperOrchestra: A Multi-Agent Framework for Automated AI Research Paper Writing A Comprehensive Implementation Guide to ModelScope for Model Search, Inference, Fine-Tuning, Evaluation, and Export
DeepSeek AI Releases DeepSeek-V4: Compressed Sparse Attention and Heavily Compressed Attention Enable One-Million-Token Contexts
Asif Razzaq · 2026-04-25 · via MarkTechPost

DeepSeek-AI has released a preview version of the DeepSeek-V4 series: two Mixture-of-Experts (MoE) language models built around one core challenge making one-million-token context windows practical and affordable at inference time.

The series consists of DeepSeek-V4-Pro, with 1.6T total parameters and 49B activated per token, and DeepSeek-V4-Flash, with 284B total parameters and 13B activated per token. Both models natively support a context length of one million tokens. DeepSeek-V4-Pro was pre-trained on 33T tokens and DeepSeek-V4-Flash on 32T tokens. Model checkpoints for all four variants: DeepSeek-V4-Pro, DeepSeek-V4-Pro-Base, DeepSeek-V4-Flash, and DeepSeek-V4-Flash-Base are publicly available on Hugging Face.

https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf

Architectural Challenges of Long Context

The vanilla attention mechanism in a standard Transformer has quadratic computational complexity with respect to sequence length, doubling the context roughly quadruples attention compute and memory. At one million tokens, this becomes prohibitive without architectural intervention. DeepSeek-V4 addresses this through four coordinated innovations: a hybrid attention architecture, a new residual connection design, a different optimizer, and FP4 quantization-aware training.

https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf

Hybrid Attention: CSA and HCA

The central architectural innovation is a hybrid mechanism combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), interleaved across Transformer layers.

CSA compresses the Key-Value (KV) cache of every m tokens into one entry using a learned token-level compressor, then applies DeepSeek Sparse Attention (DSA) where each query token attends only to the top-k selected compressed KV entries. A component called the Lightning Indexer handles sparse selection by scoring queries against compressed KV blocks. Both CSA and HCA include a sliding window attention branch covering the most recent nwin tokens for local dependency modeling.

HCA is more aggressive: it consolidates KV entries of every m′ tokens — where m′ ≫ m into a single compressed entry, then applies dense attention over those representations. No sparse selection step is needed; the compression ratio itself reduces KV cache size.

The efficiency gains are substantial. In the one-million-token setting, DeepSeek-V4-Pro requires only 27% of the single-token inference FLOPs (in equivalent FP8 FLOPs) and 10% of the KV cache size of DeepSeek-V3.2. DeepSeek-V4-Flash achieves 10% of single-token FLOPs and 7% of KV cache relative to DeepSeek-V3.2.

Manifold-Constrained Hyper-Connections (mHC)

DeepSeek-V4 replaces conventional residual connections with Manifold-Constrained Hyper-Connections (mHC). Hyper-Connections (HC) generalize residual connections by expanding the residual stream width by a factor of nhc (set to 4 in both models), introducing learned input, residual, and output mapping matrices. Naive HC suffers from numerical instability when stacking many layers.

mHC resolves this by constraining the residual mapping matrix Bl to the Birkhoff polytope — the manifold of doubly stochastic matrices where all rows and columns sum to one and all entries are non-negative. This bounds the spectral norm of the mapping at 1, preventing signal amplification in both the forward pass and backpropagation. The constraint is enforced via the Sinkhorn-Knopp algorithm with t_max = 20 iterations. Mapping parameters are dynamically generated per-input for expressivity.

Muon Optimizer and FP4 QAT

DeepSeek-V4 adopts the Muon optimizer for the majority of its parameters. Muon uses Newton-Schulz iterations to approximately orthogonalize the gradient update matrix before applying it as a weight update. The implementation uses a hybrid two-stage schedule: 8 iterations with coefficients (3.4445, −4.7750, 2.0315) for rapid convergence, then 2 stabilization iterations with coefficients (2, −1.5, 0.5). AdamW is retained for the embedding module, prediction head, static biases and gating factors of mHC modules, and all RMSNorm weights.

For deployment efficiency, FP4 (MXFP4) Quantization-Aware Training (QAT) is applied to MoE expert weights and to the Query-Key (QK) path in the Lightning Indexer of CSA. During inference and RL rollout, real FP4 weights are used directly rather than simulated quantization, reducing memory traffic and sampling latency.

Training Stability at Scale

Training trillion-parameter MoE models introduced notable instabilities. Two techniques proved effective. Anticipatory Routing decouples the backbone and routing network updates: routing indices at step t are computed using historical parameters θt−Δt, breaking the cycle in which routing decisions reinforce outlier values in MoE layers. SwiGLU Clamping constrains the linear component of SwiGLU to [−10, 10] and caps the gate component upper bound at 10, directly suppressing anomalous activations. Both techniques were applied throughout training of both models.

Post-Training: Specialist Experts and On-Policy Distillation

The post-training pipeline replaces the mixed RL stage of DeepSeek-V3.2 with On-Policy Distillation (OPD). Independent domain experts are first trained in mathematics, coding, agent tasks, and instruction following via Supervised Fine-Tuning (SFT) followed by Reinforcement Learning using Group Relative Policy Optimization (GRPO). More than ten teacher models then distill a single unified student model by minimizing the reverse KL divergence between the student and each teacher’s output distribution on the student’s own generated trajectories, using full-vocabulary logit distillation for stable gradient estimates.

The resulting model supports three reasoning effort modes: Non-think (fast, no explicit chain-of-thought), Think High (deliberate reasoning), and Think Max (maximum reasoning effort with a dedicated system prompt and reduced length penalties during RL training).

Benchmark Results

DeepSeek-V4-Pro-Max achieves a Codeforces rating of 3206, ahead of GPT-5.4-xHigh (3168) and Gemini-3.1-Pro-High (3052). On SimpleQA Verified, it scores 57.9 Pass@1, outperforming Claude Opus 4.6 Max (46.2) and GPT-5.4-xHigh (45.3), though trailing Gemini-3.1-Pro-High (75.6). On SWE-Verified, DeepSeek-V4-Pro-Max achieves 80.6% resolved, marginally behind Claude Opus 4.6 Max (80.8%), while Gemini-3.1-Pro-High also scores 80.6%.

On long-context benchmarks, DeepSeek-V4-Pro-Max scores 83.5 MMR on OpenAI MRCR 1M and 62.0 accuracy on CorpusQA 1M, surpassing Gemini-3.1-Pro-High (76.3 and 53.8 respectively), but trailing Claude Opus 4.6 Max (92.9 and 71.7) on both.

Key Takeaways

  • Hybrid CSA and HCA attention cuts KV cache to 10% of DeepSeek-V3.2 at 1M tokens.
  • Manifold-Constrained Hyper-Connections (mHC) replace residual connections for more stable deep layer training.
  • The Muon optimizer replaces AdamW for most parameters, delivering faster convergence and training stability.
  • Post-training uses On-Policy Distillation from 10+ domain experts instead of traditional mixed RL.
  • DeepSeek-V4-Flash-Base outperforms DeepSeek-V3.2-Base despite having 3x fewer activated parameters.

Check out the Paper and Model Weights. Also, feel free to follow us on Twitter and don’t forget to join our 130k+ ML SubReddit and Subscribe to our Newsletter. Wait! are you on telegram? now you can join us on telegram as well.

Need to partner with us for promoting your GitHub Repo OR Hugging Face Page OR Product Release OR Webinar etc.? Connect with us