DeepSeek V4 正式上线：1.6T 参数、100 万上下文，API 定价与性能全解析

The DeepSeek V4 series has officially opened its API. With 1.6 trillion parameters, a 1-million-token context window, and native multimodal capabilities, its Flash and Pro editions show a clear price distinction. Its coding and reasoning abilities rival the cutting-edge levels of Claude and GPT.

DeepSeek didn’t make a big splash at any launch event — it simply flipped the V4 API switch and went live.

On April 24, 2026, DeepSeek officially launched the DeepSeek V4 series models, including DeepSeek-V4-Flash and DeepSeek-V4-Pro. No teaser, no countdown — just live API endpoints. The community reacted quickly: within hours, dozens of threads appeared on Linux.do. Some exclaimed, “The big one is finally here,” while others had already started running benchmarks.

This is the largest upgrade since DeepSeek V3. The parameter count jumped from 671 billion to 1.6 trillion (1.6T); the context window expanded to 1 million tokens, with a maximum output length of 384K tokens. Together, these numbers mean you can feed an entire mid-size codebase into it, let it grasp the global context, and then refactor the code intelligently.

Two Versions, Two Positions

DeepSeek’s product strategy this time is clear: Flash for scale, Pro for quality.

| Item | DeepSeek-V4-Flash | DeepSeek-V4-Pro | |---|---|---| | Context length | 1M | 1M | | Max output | 384K | 384K | | Thinking mode | Supported (default on) | Supported (default on) | | JSON Output | ✅ | ✅ | | Tool Calls | ✅ | ✅ | | FIM Completion | Only in non-thinking mode | Only in non-thinking mode | | Input price (cache hit) | ¥0.2 / million tokens | ¥1 / million tokens | | Input price (cache miss) | ¥1 / million tokens | ¥12 / million tokens | | Output price | ¥2 / million tokens | ¥24 / million tokens |

The Flash edition sticks to DeepSeek’s usual “price butcher” style — input costs only ¥0.2 per million tokens (with cache hit), output ¥2 per million tokens. That’s roughly one-tenth of GPT-4o’s cost, making it practically negligible for daily development or batch text processing.

The Pro edition, however, is much pricier: ¥12 input and ¥24 output (when cache misses), about 12× Flash’s rate. But given the inference cost of 1.6T parameters, it’s not unreasonable. The real question: how much better is Pro than Flash? Does it justify the price delta?

Early community feedback shows Pro provides visible gains in complex reasoning and long-chain code generation, while Flash already suffices for everyday dialogue and tasks. It’s similar to Claude’s Haiku/Sonnet/Opus tiering — cheap is enough for most cases; go big when it truly matters.

So How’s the Performance?

Starting with official benchmark results: DeepSeek’s technical report includes scores covering coding, math, reasoning, and long-text comprehension. Key figures:

HumanEval score surpasses 90% — a key code generation benchmark
SWE-Bench test reaches 84% — strong performance on real software engineering tasks
Inference speed ~2× faster than V3

Of course, you should take official scores with a grain of salt. As of publication, third-party benchmark platforms such as LMArena (ex Chatbot Arena), ArtificialAnalysis, and LiveBench have not yet released V4 data. True battles will emerge when community blind tests roll out.

Still, available information suggests V4 has solid progress in several areas:

1. Long-context capability.
A 1 million-token window isn’t for show — V3 had 128K tokens, so this upgrade is ~8×. That’s a qualitative leap for large codebase work, long document analysis, and complex multi-turn dialogue. Tasks that previously required RAG retrieval can now be handled directly in context.

2. Coding ability.
Internal test results cited by multiple outlets indicate V4 might outperform Claude and the GPT series in programming tasks. Might — pending third-party blind verification. But a 90%+ HumanEval score speaks volumes, ranking V4 among the best public models today.

3. Multimodality.
V4 natively supports image, text, and video input — a capability V3 didn’t have. Although current API access is mainly text-based, the multimodal foundation is already laid.

Architectural Innovations

DeepSeek V4 continues using an MoE (Mixture of Experts) architecture, but introduces an “Engram memory technology.” In short: MoE activates only part of the parameters per inference (so 1.6T parameters ≠ running all of them each time), while Engram enhances memory and retrieval efficiency under long-context scenarios.

That explains why V4 can double its parameters yet speed up inference by 2× — not through brute force hardware, but architectural efficiency.

Another key detail: DeepSeek V4 is optimized for domestic AI chips such as Huawei Ascend, offering early access and tuning with local hardware vendors before release — but no pre-release provided to NVIDIA or AMD. The signal is clear: amid export control challenges, DeepSeek is proactively reducing dependency on the NVIDIA ecosystem.

For developers, this means local deployment on Chinese hardware should be smoother. The model weights are available for download on Hugging Face (deepseek-ai/DeepSeek-V4 Collection), but local setup for a 1.6T-parameter model is challenging — most developers will realistically use the API.

Impact on the Developer Ecosystem

The community is already asking a practical question: How soon will domestic Token Plan and Code Plan platforms integrate V4?

The reason: numerous domestic API-based tools — code-completion plugins, AI coding assistants, workflow automation platforms — depend on which model they use, affecting user experience and cost directly. V4’s arrival gives these providers a strong new option.

The Flash version deserves special attention. With ¥1/million tokens input (cache miss), ¥2/million tokens output, a 1M context window, and solid coding competence — this price-performance mix is ideal for AI coding tools like Cursor, Continue, and OpenCode.

Some community members predict overseas aggregators such as OpenRouter will integrate V4 soon. Domestic ones are following too — platforms supporting OpenAI-compatible formats (like OpenAI Hub) generally add new models quickly, so developers should stay tuned.

However, skeptics note: a 1.6T model comes with nontrivial inference cost. Pro pricing already hints at it — running full-power V4 demands compute muscle. For smaller API providers, managing concurrency and cost control will be challenging.

Where V4 Stands Among Competitors

Placed in the current landscape:

vs Claude 3.5/4: V4 Pro may already match or slightly surpass Claude on coding tasks (to be verified), though Claude’s instruction-following and text stability still set the standard. Price-wise, V4 Flash is far cheaper, and even Pro undercuts Claude.
vs GPT-4o/GPT-5: V4’s 1M context length (vs GPT’s 128K–256K) is a clear win. In coding and reasoning, V4 Pro competes toe-to-toe with GPT-4o; the gap vs GPT-5 remains to be seen.
vs Domestic Models (Qwen, Kimi, MiniMax, GLM): V4’s parameter and context lengths raise the bar significantly. Some YouTube reviews even titled videos “Completely crushes other Chinese models” — exaggerated wording aside, V4 currently sits at the top domestically. Kimi used to lead with long context, but V4’s 1M window and expanded parameters neutralize that edge.

An interesting observation: V4 offers both OpenAI-format and Anthropic-format API endpoints.

# OpenAI-compatible format
BASE_URL = "https://api.deepseek.com"

# Anthropic-compatible format
BASE_URL = "https://api.deepseek.com/anthropic"

This means that whether your project is wired for OpenAI or Claude, migration to V4 costs almost nothing — just change base_url and model name. A smart compatibility move: lowering the migration barrier equals user acquisition.

Key Details Worth Noting

Thinking mode enabled by default.
Both versions support “thinking” and “non-thinking” modes, with thinking mode on by default. This continues DeepSeek R1’s philosophy — “think first, answer later,” boosting accuracy on complex tasks. For latency-sensitive or simple jobs, manually disable thinking mode.
FIM (Fill-in-the-Middle) completion only works in non-thinking mode.
This matters for code autocompletion — if using V4 inside an IDE, disable thinking mode, or the FIM interface won’t be active.
Conversation prefix continuation in Beta.
This allows specifying a prefix and having the model generate from it — handy when precise output formatting is required.
Maximum output of 384K tokens.
That’s massive — roughly equal to a 500,000-word book. You’ll rarely need such a long single output, but it provides ample space for ultralong reasoning chains or extensive code generation tasks.

Keep a Cool Head

After all the hype, let’s temper expectations.

First, official scores never equal real-world experience. A 90%+ HumanEval is impressive, but those problems are simple; real coding is messier. SWE-Bench at 84% is more representative but still dataset-dependent.

Second, a 1.6T model may have stability issues — more parameters don’t guarantee better output every time. Sometimes large models “overthink” and fare worse on simple queries.

Third, third-party test platforms (LMArena, LiveBench, etc.) haven’t published V4 ranks yet. Claims of “beating Claude” or “crushing GPT” remain anecdotal until blind tests drop.

Finally, while overall cheaper than overseas models, Pro’s ¥12/million tokens input (cache miss) isn’t cheap for frequent calls. The good news: cache hit drops it to ¥1 — so smart caching matters.

Summary

The launch of DeepSeek V4 marks a major leap for domestic LLMs — in parameter scale, context length, and multimodal capability. The Flash edition’s extreme cost-performance makes it a likely daily driver for developers, while the Pro edition targets high-precision professional use cases.

Developers can already request an API key at api.deepseek.com. If your tooling supports OpenAI-compatible format, migration is almost effortless.

But don’t rush to call it “the strongest model.” Wait for third-party blind tests and community feedback over the next weeks. Benchmarks are numbers — usability is how it feels.

References: