DeepSeek Launches V4 Preview, a 1M Context Open Source Model Built to Rival Closed-Source AI
Chinese AI startup DeepSeek released a preview version of its long-awaited V4 large language model on Friday, allowing developers to test its capabilities ahead of a full launch. The release comes more than a year after DeepSeek's R1 reasoning model disrupted global tech markets with performance that matched leading U.S. models at a fraction of the reported cost.
V4 Comes in Two Versions Built for Different Use Cases
V4 launches as two variants built on a Mixture-of-Experts architecture:
- DeepSeek-V4-Pro: 1.6 trillion total parameters, 49 billion active. Targets complex reasoning, agentic coding, and world knowledge tasks.
- DeepSeek-V4-Flash: 284 billion total parameters, 13 billion active. Matches Pro in reasoning for simple tasks and agent performance while offering faster and lower-cost API access.
Both versions support a 1 million token context window, enabled by a new DSA sparse attention mechanism that reduces computational and memory demands for long context processing. Both are open source with model weights available on Hugging Face.
The existing API model names deepseek-chat and deepseek-reasoner will be retired on July 24, 2026.
How V4-Pro Stacks Up Against the Competition
According to DeepSeek, V4-Pro:
- Achieves the best open-source level in agentic coding evaluations, with delivery quality close to Claude Opus 4.6
- Surpasses all publicly evaluated open-source models in math, STEM, and competition code
- Significantly leads other open-source models in world knowledge benchmarks, falling only slightly behind Google's Gemini 3.1 Pro
The model supports both a non-thinking mode and a thinking mode with an adjustable reasoning effort parameter in the API, and has been optimized for popular agent tools, including Anthropic's Claude Code, OpenClaw, OpenCode, and CodeBuddy.
Why This Launch Comes at a Critical Moment for DeepSeek
DeepSeek gained global attention in late 2024 with V3, which it said was trained on less powerful chips at a fraction of the cost of models from OpenAI and Google. In January 2025, R1 matched or outperformed many leading LLMs on key benchmarks, raising immediate questions in tech markets about the scale of spending required for AI infrastructure.
Since R1, no DeepSeek release has matched its market impact.
V4 arrives amid a sharper competitive and financial backdrop:
- Alibaba and ByteDance have both released new models in 2026, intensifying competition inside China
- Reuters reported that DeepSeek gave Huawei's Ascend 950PR chips exclusive early hardware access to V4 while denying Nvidia early access, a deliberate move amid ongoing U.S. export controls on advanced semiconductors bound for China
- Tencent and Alibaba are competing to invest in DeepSeek's first external fundraising round, targeting a valuation of over $20 billion, according to aibase
"This addresses the long-standing issues of slower performance and higher costs associated with long context lengths, marking a genuine inflection point for the industry," Zhang Yi, the founder of tech research firm iiMedia, told AFP. "For end users, this will bring widespread, accessible benefits. For instance, if ultra-long context support becomes a standard feature, long-text processing is expected to move beyond high-end research labs and enter mainstream commercial applications," he said.
Also Read
- DeepSeek V3.1 Hybrid Reasoning Model Released
- Google Launches Gemma 4, Its Most Capable Open Model Family
- Moonshot AI Open-Sources Kimi K2.6, a Coding Model That Runs Autonomously for Days
- Microsoft Copilot Adds Multi-Model AI Comparison
- OpenAI Closes Record $122 Billion Funding Round at $852 Billion Valuation
