DeepSeek-V4 Preview

Overview

DeepSeek-V4 is the April 24, 2026 Preview release of DeepSeek's flagship model family, launched simultaneously in two variants: V4-Pro (1.6 trillion total / 49 billion active parameters) and V4-Flash (284 billion total / 13 billion active). Both variants ship with a 1M-token context window as the default, a new Hybrid Attention Architecture combining token-wise compression with DeepSeek Sparse Attention (DSA), and open weights on Hugging Face under the MIT License. V4 succeeds DeepSeek-V3.2 as the mainline general-purpose model, with OpenAI ChatCompletions and Anthropic API wire compatibility that lets developers migrate existing client code with only a base-URL change. The release positions DeepSeek as a leading open-weight alternative for agentic coding and long-context workflows, with V4-Flash and V4-Pro priced well below GPT-5.5 and Claude Opus 4.7 standard API rates.

What's New

Two Variants Released Simultaneously

V4 ships as two distinct models rather than a single flagship:

• DeepSeek-V4-Pro: 1.6T total parameters with 49B activated per token. Targets the hardest reasoning, math, agentic coding, and research workloads where capability justifies the higher price.
• DeepSeek-V4-Flash: 284B total / 13B active parameters. Delivers reasoning that "closely approaches V4-Pro" per DeepSeek's release notes, runs significantly faster, and matches V4-Pro on simpler agentic tasks at roughly one-twelfth the API price.

Both variants share the same architecture, tokenizer, and 1M-token default context window. The choice between them is primarily about cost-per-task and latency rather than capability tier.

Hybrid Attention Architecture + DSA

V4 introduces a Hybrid Attention Architecture combining Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA), building on the sparse-selection ideas from DeepSeek Sparse Attention in the V3.2 line. The architecture delivers:

• 1M token context as the default across official DeepSeek services, API, and open weights, without a separate long-context tier.
• Lower long-context compute and memory cost: in the 1M-token setting, V4-Pro uses 27% of V3.2's single-token inference FLOPs and 10% of its KV cache; V4-Flash uses 10% and 7%, respectively.
• Substantial efficiency gains in both training and inference on long-context scenarios, maintaining output quality of dense attention while running at fraction of compute cost.

Benchmark Performance

DeepSeek positions V4-Pro as the first open-weight model to credibly compete with closed-source frontier models on the hardest evaluations:

• Math, STEM, and coding: V4-Pro-Max posts strong official results across published benchmarks, including GPQA Diamond 90.1, LiveCodeBench 93.5, Terminal Bench 2.0 67.9, SWE Verified 80.6, and SWE Pro 55.4.
• Agentic Coding: Open-source state-of-the-art on agentic coding benchmarks per DeepSeek's release note. The official model card publishes head-to-head rows against GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro, K2.6 Thinking, and GLM-5.1 Thinking.
• World Knowledge: DeepSeek says V4-Pro leads current open models and trails only Gemini 3.1 Pro among the frontier models shown in its comparison; Gemini 3.1 Pro itself is not an open-weight model.
• V4-Flash: "Closely approaches V4-Pro" on reasoning and matches V4-Pro on simple agent tasks, per DeepSeek's own internal evaluations.

Numeric scores are published on the official Hugging Face model card for MMLU-Pro, GPQA Diamond, LiveCodeBench, SWE Verified, SWE Pro, Terminal Bench 2.0, BrowseComp, and related benchmarks; the full tech report is expected to follow with deeper architectural detail.

Dual Modes + Agent Framework Integration

Both V4-Pro and V4-Flash support reasoning-mode controls. The model card lists three reasoning effort modes—Non-think, Think, and Think Max—while the API docs describe switching between non-thinking and thinking modes:

• Thinking mode — Explicit chain-of-thought reasoning for harder problems where accuracy matters more than latency.
• Non-Thinking mode — Direct answer generation for simpler tasks where latency matters more than extended reasoning.

Integration with leading agent frameworks including Claude Code and OpenClaw is supported natively — V4 is positioned as a low-cost drop-in for teams already running AI agent workflows on paid closed models.

Legacy Model Deprecation Schedule

DeepSeek's OpenAI-style legacy endpoints deepseek-chat and deepseek-reasoner are scheduled to retire on July 24, 2026, three months after the V4 Preview launch. Existing integrations should plan migration to deepseek-v4-flash or deepseek-v4-pro endpoints before that date. The V4 endpoints accept the same request format, so migration is typically a model-name change rather than a protocol rewrite.

Availability & Access

Launch Availability (Apr 24, 2026)

V4-Pro and V4-Flash rolled out simultaneously across three surfaces on April 24, 2026:

• chat.deepseek.com: Free access with Expert Mode (Thinking) and Instant Mode (Non-Thinking) both defaulting to V4.
• API (platform.deepseek.com): Pay-as-you-go pricing available day one, OpenAI ChatCompletions and Anthropic API formats both supported.
• Open Weights (Hugging Face): deepseek-ai/DeepSeek-V4-Pro, deepseek-ai/DeepSeek-V4-Flash, and their base checkpoints are published under the MIT License, with local deployment instructions in DeepSeek's official inference repository.

A technical report PDF accompanies the release, detailing the Hybrid Attention Architecture and DSA design choices for research teams wanting to reproduce or audit the work.

Regional & Compliance Considerations

The DeepSeek-hosted API and web surface are subject to Chinese jurisdiction and the provider's stated retention policy. Teams with strict data-residency, geopolitical-risk, or regulated-data requirements should evaluate the self-hosted open-weight path, which removes provider-hosted-traffic questions entirely. The open weights are published under the MIT License, which broadly permits commercial use, modification, distribution, and self-hosting subject to the license terms.

Pricing & Plans

V4 uses a three-tier access model identical to prior DeepSeek generations: free web chat, pay-as-you-go API, and free open weights.

chat.deepseek.com (Web)

• Free web access — V4 is available at chat.deepseek.com via Expert Mode and Instant Mode. Specific usage limits and account requirements are at DeepSeek's discretion.

API — Pay As You Go

Both endpoints support OpenAI ChatCompletions and Anthropic message formats, and both support Thinking / Non-Thinking dual modes.

Comparison with closed-source flagships:

V4-Pro lands at roughly one-third the input price and one-ninth the output price of GPT-5.5. V4-Flash pricing is more than an order of magnitude below GPT-5.5 and Claude Opus 4.7 standard API rates, making it attractive for high-volume workloads where cost dominates the cost-quality tradeoff.

Open Weights (Self-Hosted)

• Free under the MIT License from Hugging Face.
• V4-Flash (284B / 13B active) is the more practical self-hosting choice; V4-Pro's 1.6T parameters require substantial GPU infrastructure.
• vLLM and SGLang supported out of the box; quantized community builds typically appear within days of release.

Best For

• Cost-sensitive developers and startups moving off GPT-5.5 or Claude Opus 4.7 for high-volume workloads where V4-Flash's 35× cost advantage compounds quickly.
• Agentic coding teams using Claude Code, OpenClaw, or similar harnesses who want an open-weight drop-in at one-third the token cost.
• Long-context and codebase-wide workflows (entire repos, multi-document analysis, extended research) that benefit from 1M context without premium tiering.
• Research, math, and STEM users where V4-Pro's published math, STEM, coding, and reasoning benchmark results indicate strong capability.
• Self-hosters and privacy-sensitive organizations needing frontier capability on air-gapped or on-premise infrastructure.
• Teams maintaining DeepSeek deepseek-chat or deepseek-reasoner integrations who need to migrate before the July 24, 2026 legacy retirement deadline.

FAQ