Chinese AI startup DeepSeek has released a preview version of its V4 model series, introducing a one-million-token context window and enhanced agent capabilities that approach leading closed-source competitors, while maintaining its commitment to open-source accessibility.

The company launched two preview variants on April 24, 2026: DeepSeek-V4-Pro, designed to match proprietary models in reasoning and world knowledge, and DeepSeek-V4-Flash, optimized for speed and cost efficiency. Both models are immediately accessible through DeepSeek's API and chat interface at chat.deepseek.com, with preview weights released on Hugging Face and ModelScope.

Agent Performance Narrows Gap With Established Players

DeepSeek-V4-Pro's preview demonstrates substantial improvements in agentic coding tasks, achieving what the company characterizes as best-in-class performance among open-source models. Internal testing by DeepSeek employees indicates the preview model's coding experience exceeds Anthropic's Claude Sonnet 4.5 and approaches Claude Opus 4.6's non-reasoning mode, though still trails the reasoning-enabled Opus variant.

The preview has been specifically tuned for integration with mainstream agent frameworks including Claude Code, OpenClaw, OpenCode, and CodeBuddy, showing measurable gains in code generation and document creation workflows. DeepSeek reports improved output quality in complex multi-step operations, positioning the V4 preview as a potential alternative for enterprise developer tools pending full release.

Novel Architecture Delivers Context Efficiency

The V4 preview introduces an experimental attention mechanism that compresses information at the token dimension while utilizing DeepSeek Sparse Attention (DSA). This architectural approach enables the preview model to process up to one million tokens—approximately 750,000 words—while significantly reducing computational requirements compared to traditional transformer implementations.

The efficiency advantages become pronounced at extended context lengths. According to DeepSeek's technical documentation, the V4 preview maintains substantially lower memory overhead and compute demands than its predecessor V3.2 when processing contexts beyond 100,000 tokens. The company positions the million-token capacity as a standard feature rather than premium capability, though as a preview release, performance characteristics may evolve before final deployment.

Benchmark Results Show Competitive Preview Performance

In world knowledge evaluations, the V4-Pro preview substantially outperforms other open-source models, trailing only Google's Gemini-Pro-3.1 among all tested systems. The preview demonstrates particular strength in STEM reasoning, mathematics, and competitive programming benchmarks, matching or exceeding results from established closed-source alternatives in current testing.

DeepSeek-V4-Flash preview, while showing slightly reduced world knowledge retention due to its smaller parameter count and activation footprint, maintains comparable reasoning capabilities to the Pro variant on most tasks. The Flash preview performs equivalently to Pro on straightforward agent operations but shows measurable gaps on high-complexity scenarios requiring extended reasoning chains.

API Access Marks Preview Deployment Phase

The preview API release introduces a three-month transition timeline for legacy model endpoints. The existing deepseek-chat and deepseek-reasoner endpoints will cease operation on July 24, 2026, redirecting to V4-Flash preview's non-reasoning and reasoning modes respectively. Developers must update model parameters to deepseek-v4-pro or deepseek-v4-flash while maintaining existing base URLs.

Both V4 preview variants support OpenAI ChatCompletions and Anthropic API interfaces, with reasoning mode accessible through a reasoning_effort parameter offering "high" or "max" intensity settings. DeepSeek recommends maximum reasoning intensity for complex agent workflows in the preview phase, though this increases inference latency and token consumption during evaluation.

The company maintains its established pricing structure for the preview period, with V4-Flash positioned as a cost-optimized option compared to V4-Pro. Exact API pricing was not disclosed in the announcement, though DeepSeek historically prices below Western closed-source providers while matching capability levels. Preview pricing may adjust upon full release.

Open-Source Preview Continues Transparency Strategy

Preview model weights and technical documentation are available through Hugging Face and China's ModelScope platform, continuing DeepSeek's practice of releasing production-stage models without restrictions. The accompanying technical report details the experimental sparse attention architecture and training methodology, enabling academic and commercial evaluation during the preview phase.

The open-source preview approach contrasts sharply with Western AI leaders' increasing reluctance to publish model weights. DeepSeek's willingness to release state-of-the-art preview capabilities could accelerate global AI development while complicating US export control strategies targeting Chinese AI advancement.

The V4 preview launch arrives amid heightened geopolitical scrutiny of Chinese AI companies. DeepSeek's earlier V3 model triggered market volatility in January 2026 when its efficiency metrics suggested Chinese firms could match Western capabilities with fewer resources. The V4 preview's emphasis on architectural innovation and open access reinforces that narrative, potentially influencing both investment patterns and regulatory frameworks as the model progresses toward full release.

Related Coverage：

DeepSeek Overhauls GPU Kernels to Slash AI Compute Overhead

DeepSeek Seeks $300M at $10B Valuation as Nvidia Dependence Becomes Existential Liability

DeepSeek V4 Targets Late April Launch, Betting on Trillion-Parameter Efficiency