On April 24, 2026, DeepSeek officially released and open-sourced the DeepSeek-V4 series. This 1.6T-parameter open-source behemoth upgrades across the board in Agent coding, long context, and complex reasoning, and shot to the top of Hugging Face's trending list on launch day. This article rounds up the core specs, API pricing, and multi-angle test results to help you quickly judge whether it's worth switching.

Core Specs

The DeepSeek-V4 series includes two versions, both built on a MoE (Mixture of Experts) architecture:

In the 1 million-token scenario, V4's per-token inference compute is only 27% of V3.2's, and KV Cache usage drops to about 10%.

API Pricing

V4-Flash's pricing is extremely aggressive — RMB 0.2 per million tokens on cache hits, well suited to lightweight, high-frequency calls.

Note: DeepSeek has stated that, constrained by high-end compute, V4-Pro's serving throughput is currently limited. Prices are expected to drop substantially once the Ascend 950 supernode ships in volume in the second half of the year.

Hands-On: 60 Minutes of Continuous Coding with No Intervention

Task 1: Building a complete bookkeeping system from scratch

Testers gave DeepSeek-V4-Pro a vague database design requirement. The model first autonomously planned a solution with 8 functional modules and 6 database tables, then laid out a 10-item verification checklist.

Once planning was done, they let it run completely hands-off: nearly 60 minutes of continuous coding — no interruptions, no infinite loops, no missed critical steps. Every core frontend feature worked, and the front-end/back-end data flow was complete end to end.

This task consumed about 200,000 tokens (mostly input), with an API bill of roughly RMB 5. With the caching mechanism kicking in, the actual cost stayed under control.

Task 2: Building an Android emulator from scratch

The complexity of this task exceeded the model's knowledge, but it autonomously started web searches, made 18 tool calls, and after 11 minutes of thinking produced a complete seven-step development plan. It then ran continuously for over 50 minutes to finish the job.

The key takeaway: when the model hits a knowledge blind spot, it proactively searches to fill the gap instead of making things up.

Long-text test: reading the entire Three-Body trilogy in one go

Uploading the full text of The Three-Body Problem trilogy consumed 540,000 tokens. The model accurately answered questions about plot details from the books, validating that the million-token context is genuinely usable.

Known Weaknesses

• Occasional stumbles on lightweight tasks: On simple questions the Pro version sometimes "overthinks", while the Flash version is more direct and efficient
• Reasoning stability needs improvement: Infinite loops appear in some extreme math problems and lightweight tests
• Prices are up from V3.2: But the caching mechanism partially offsets the increase

Quick Start

• Web experience: chat.deepseek.com
• API docs: api-docs.deepseek.com
• Open weights: HuggingFace / ModelScope
• Technical report: DeepSeek_V4.pdf