DeepSeek V4 Release Date & Features | Native Multimodal, 1T Params, March 2026

DeepSeek V4 is natively multimodal — trained on text, image, video and audio from scratch. Unlike competitors that bolt vision onto text models, V4 understands all modalities natively.

• • Processes text, image, video, audio natively
• • Trained on multimodal data from scratch
• • Not a text model with bolted-on vision
• • Unified understanding across all modalities

V4 features 1 trillion total parameters with only 32B active per token via Mixture-of-Experts. This delivers frontier performance at 10-25x lower cost than GPT-5.4.

• • 1T total parameters, 32B active per token
• • Mixture-of-Experts (MoE) architecture
• • API pricing: $0.10-$0.30 per million tokens
• • 10-25x cheaper than GPT-5.4, open-source

Expected to support million-level token context window, can process entire books, large codebases or ultra-long documents.

• • Expand from current 128K to million-level
• • Support processing entire books (~500K words)
• • Can analyze complete large project codebases
• • Multi-turn conversation memory greatly enhanced

Revolutionary conditional memory mechanism enabling effectively infinite context. Retrieves relevant memories in O(1) time, allowing V4 to recall your entire codebase or knowledge base instantly.

• • O(1) memory retrieval for instant recall
• • Effectively infinite context window
• • Recall entire codebases and knowledge bases
• • Conditional memory replaces traditional KV Cache

Novel sparse attention mechanism that reduces computational costs by ~50% while supporting context windows exceeding 1 million tokens. Combined with FP8 mixed precision for maximum efficiency.

• • Computational cost reduced ~50%
• • Enables 1M+ token context windows
• • FP8+bfloat16 mixed precision inference
• • Memory usage reduced 50%+ via FP8 KV Cache

Features a 'pause and think' Chain-of-Thought mechanism similar to OpenAI o1. V4 can break down complex problems, reason step-by-step, and self-correct before outputting answers.

• • Chain-of-Thought 'pause and think' mechanism
• • Multi-step reasoning for complex problems
• • Self-correction before final output
• • 40% jump in reasoning benchmarks over V3

DeepSeek V4 API pricing projected at $0.10-$0.30/M tokens. GPT-5.4 costs $2.50-$15/M. Cache hits reduce cost by 90%. Open-source means free self-hosting.

• • Input: $0.10-$0.30 per million tokens
• • Cache hit: 90% discount on input
• • 10-25x cheaper than GPT-5.4 ($2.50-$15/M)
• • Open-source: free to self-host

Internal benchmarks target 80%+ on SWE-bench Verified, competing with Claude 4.6 (80.8%), Gemini 3.1 Pro (80.6%), and outperforming GPT-5.4 (77.2%) — at 10-80x lower cost.

• • SWE-bench target: 80%+ (vs Claude 4.6's 80.8%, Gemini 3.1's 80.6%)
• • HumanEval coding: 90%+ expected
• • Outperforms GPT-5.4 (77.2%) at 10-25x lower cost
• • 50+ language support, repo-level bug fixing