DeepSeek-V4: a million-token context that agents can actually use

2026-04-24 · Source: Hugging Face - Blog · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, medium

Summary

DeepSeek has released V4, featuring two Mixture-of-Experts (MoE) checkpoints: DeepSeek-V4-Pro (1.6T total parameters, 49B active) and DeepSeek-V4-Flash (284B total, 13B active), both offering a 1M-token context window. The models prioritize efficient long-context support for agentic tasks, addressing common failures like context budget overruns and KV cache exhaustion. DeepSeek-V4 achieves significant efficiency gains, requiring 27% of single-token inference FLOPs and 10% of KV cache memory compared to DeepSeek-V3.2 for V4-Pro, and even lower for V4-Flash. This efficiency stems from a hybrid attention architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), interleaved across layers, and utilizing FP8/FP4 storage. Additionally, V4 preserves reasoning across user message boundaries for tool-use conversations, introduces an XML-based tool-call schema with a "|DSML|" token, and was trained using DSec, a Rust-based sandbox for RL rollouts.

Key takeaway

For AI Architects and MLOps Engineers deploying agentic systems, DeepSeek V4 offers a compelling solution to long-context challenges. Its optimized architecture and agent-specific training, including persistent reasoning and a robust tool-call schema, directly address common failure modes in multi-step agent workflows. Consider evaluating DeepSeek V4-Pro or V4-Flash for your next agent-powered application, especially where context length and inference efficiency are critical.

Key insights

DeepSeek V4 optimizes large language models for agentic workflows through architectural and training innovations.

Principles

• Efficiency enables utility for long contexts.
• Hybrid attention reduces KV cache and FLOPs.
• Preserving reasoning improves agent coherence.

Method

DeepSeek V4 employs a hybrid attention mechanism (CSA and HCA) with interleaved layers and FP8/FP4 storage for efficiency. It preserves reasoning across tool-call turns and uses an XML-based tool-call schema.

In practice

• Use DeepSeek V4 for long-running agentic tasks.
• Implement "|DSML|" XML schema for robust tool calls.
• Leverage Think Max mode for complex reasoning.

Topics

• DeepSeek-V4
• Agentic Workloads
• Long-Context Inference
• Hybrid Attention
• KV Cache Optimization

Best for: AI Architect, MLOps Engineer, CTO, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• As agentic AI pushes rivals to raise prices and cap usage, Deepseek ships a good-enough model for almost nothing — Deepseek's V4 models offer massive context windows and competitive performance at significantly reduced costs via architectural innovation.
• 🥇Top AI Papers of the Week — DeepSeek V4 pushes open-source LLMs to million-token contexts with novel attention and training stability techniques.
• Build with DeepSeek V4 Using NVIDIA Blackwell and GPU-Accelerated Endpoints — DeepSeek V4 models optimize long-context inference via a hybrid attention mechanism, significantly reducing FLOPs and KV cache memory.
• DeepSeek-V4: Towards Highly Efficient Million-Token Context Intelligence — The DeepSeek-V4 series achieves highly efficient million-token context processing through architectural innovations and optimized training.
• DeepSeek-V4 Pro now available on Together AI — DeepSeek-V4 Pro offers scalable, cost-effective long-context reasoning with flexible deployment and controllable reasoning modes.
• DeepSeek V4: One Million Tokens, Three Thinking Modes, and the First Real Hands-On Reports — DeepSeek-V4 optimizes long-context LLMs via novel attention and training stability, offering tiered reasoning and cost-effective variants.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Hugging Face - Blog.