DeepSeek API maxes out at a million tokens overnight — the V4 era arrives early

From 128K to 1M — DeepSeek API Fully Unleashed Today

On April 22, many developers noticed that the official DeepSeek API’s context window had jumped from 128K tokens to 1 million tokens—fully aligning with the previously gray‑tested App and web versions. This isn’t a minor iteration; it’s a leap in magnitude.

Based on previously leaked information, this API version is very likely DeepSeek V4 Lite, the lightweight variant of the V4 series designed for API scenarios. Liang Wenfeng previously stated that V4 would launch in late April; it now appears that the Lite version has arrived early.

What Does One Million Tokens Mean

Let’s start with an intuitive comparison.

128K tokens can handle a Chinese book of roughly 100,000 characters—enough for daily dialogues and short document analysis. But when it comes to real engineering use cases—like reviewing an entire codebase, analyzing dozens of contracts in one go, or having the model read a full novel before summarizing—128K quickly runs out of room.

So what does 1 million tokens mean? Roughly 750,000 Chinese characters, about the length of four volumes of The Three Body Problem. You can stuff an entire mid‑sized project’s source code, documentation, and test cases into a single request, allowing the model to reason over the full context instead of relying on RAG retrieval to piece together fragments.

This is a tangible benefit for several types of developers:

• Long‑document processing: legal review, financial reports, academic literature surveys—no more splitting into chunks
• Code analysis: repo‑level understanding, refactoring suggestions, security audits—the context is finally large enough
• Agent applications: longer conversation memory without frequent truncation—Agent “memory” maxed out
• RAG workflows: with large enough context windows, many cases that used to require complex retrieval pipelines can simply feed everything into the prompt

Of course, a bigger window doesn’t automatically mean better results. There’s a long‑standing industry consensus: many models claim ultralong context support, but their retrieval and reasoning performance drops significantly in the latter half of the window—the classic “Lost in the Middle” problem. How DeepSeek V4 Lite performs here remains to be proven by community benchmarks. Still, initial developer feedback suggests it’s usable for long‑text scenarios.

Knowledge Base Updated to May 2025: A Change Easy to Miss

Beyond the window upgrade, there’s another noteworthy change: the knowledge cutoff has been updated to May 2025.

This means the model can accurately answer questions about events up to April 2025, even without internet access—good enough for many offline or isolated deployments.

For comparison:

| Model | Knowledge Cutoff | Context Window | |--------|------------------|----------------| | DeepSeek V3.1 (old) | Feb 2025 | 128K | | DeepSeek V4 Lite (new) | May 2025 | 1M | | GPT‑4o | Oct 2024 | 128K | | Claude 3.5 Sonnet | Apr 2025 | 200K | | Gemini 2.5 Pro | Jan 2025 | 1M |

In context length, DeepSeek V4 Lite is now on par with Google Gemini 2.5 Pro’s 1M and far ahead of OpenAI and Anthropic’s current flagships. Its knowledge freshness also ranks among the top tier.

However, one detail is clear: V4 Lite currently does not support image input, only text and speech. Although leaks show V4 Lite possesses native multimodal capability (image understanding and OCR included), today’s API update doesn’t expose vision features yet. This might be a staged rollout—or they’re holding it until the full V4 launch.

V4 Lite’s Positioning: Not a Cut‑Down Version, but a Sharp Tool

Judging by its name and known details, V4 Lite’s place in the V4 lineup is clear: low latency, fast response, API‑first.

This mirrors OpenAI’s GPT‑4o mini and Anthropic’s Claude 3.5 Haiku—the same philosophy: not every scenario needs maximum reasoning power; often developers want something fast, inexpensive, and good enough. The Lite edition is built for exactly that.

From today’s real‑world developer experience, V4 Lite is noticeably faster than the V3 series, especially on long‑text inputs—the first‑token time (TTFT) improves significantly. For production environments processing large text volumes, that’s crucial—you don’t want to feed it a book and wait 30 seconds for the first word to appear.

As for the flagship V4, per Liang Wenfeng’s earlier remarks, it’s also expected in late April. According to leaks, the flagship has 671 billion parameters and will be a truly native multimodal model supporting image understanding and generation. If the Lite version is the appetizer, the flagship will be the main course.

An Unusually Busy Month for DeepSeek

Looking at the timeline, DeepSeek’s pace over the past month has been intense:

• Apr 8 – Expert mode launch for deeper reasoning on complex problems
• Mid‑Apr – Expert mode gains file upload support
• Mid‑Apr – Million‑token context gray‑tested on app & web
• Apr 22 (today) – API upgraded to 1M tokens; V4 Lite released
• Late Apr (expected) – V4 flagship official launch

There was also a three‑day service outage affecting both web and API. Given the heavy infrastructure changes and model swaps, such instability isn’t surprising—but for developers relying on DeepSeek API in production, it’s a risk worth noting.

This highlights a practical issue: availability risk from a single API provider. If your service depends on one model vendor, any outage directly impacts your business. Hence many developers are adopting multi‑model aggregation—connecting through a unified API gateway to multiple models, enabling quick failover and cost‑based model selection across scenarios. Aggregator platforms like OpenAI Hub already support DeepSeek’s new million‑token API, sparing developers the pain of managing multiple API keys for GPT, Claude, Gemini, and DeepSeek.

How the Million‑Token Era Changes Development Patterns

As mainstream models reach million‑token windows, traditional development patterns are being reevaluated.

RAG’s role is shifting. When context windows were just 4K or 8K, RAG (Retrieval‑Augmented Generation) was almost mandatory for long documents: split, vectorize, store, then retrieve snippets to insert into the prompt. Effective—but complex and lossy.

Now, with 1M tokens, many use cases can simply feed raw documents directly, skipping retrieval. This doesn’t kill RAG—once datasets grow to GBs, full‑context loading is unrealistic—but for small‑to‑medium collections, brute‑force context often beats finely tuned RAG pipelines because the model sees the whole story and misses no key facts.

Agent capability ceiling rises. Long context lets Agents retain more conversation history and task state without frequent compression or summarization. An Agent that “remembers” dozens of prior turns behaves far more coherently in complex tasks.

Cost structure must be recalculated. A million‑token input means potentially 8× the token use per request. DeepSeek’s pricing is aggressive, but heavy API usage still piles up fast. Developers need to balance “more context for better results” against “token cost control.”

Practical tip: Don’t stuff everything in just because you can. Assess how much context your scenario really needs before choosing window size. A well‑organized 100K input often outperforms a careless 1M dump.

Competitive Landscape: Million Tokens Becoming the Default

Across the industry, million‑token windows are turning from “differentiator” into “baseline”:

• Google Gemini 2.5 Pro first brought 1M to the API layer
• DeepSeek V4 Lite joins the club today
• Anthropic’s Claude still caps at 200K but likely extending soon
• OpenAI’s GPT series seems conservative here, holding 128K for some time

Interestingly, in this “context‑window arms race,” Chinese model vendors are not lagging. DeepSeek’s Lite matches Gemini’s window size—and given DeepSeek’s usual pricing, likely wins on cost‑effectiveness.

Of course, window size is only one dimension. Reasoning quality, instruction following, code generation, and multimodal understanding are what truly matter for production use. V4 Lite’s performance on these fronts still needs systematic evaluation over time.

Final Thoughts

Today’s API update feels like a warm‑up before V4’s official release—the Lite edition first, the flagship next. DeepSeek’s roadmap is clear.

For developers, now’s the time to test million‑token contexts in real scenarios, starting with your most painful long‑text cases—those that previously required hacks to work around context limits. Try the large window directly and see whether results and costs justify it.

The flagship V4 will likely drop in the coming days. Only with its full benchmarks and pricing can we truly assess DeepSeek’s competitiveness this generation.

---

References

• ITHome: DeepSeek API adds 1‑million‑token context, matches App and Web versions — main news source with first‑hand user feedback
• Zhihu: DeepSeek quietly tests new model — million‑token context and updated knowledge base — detailed analysis from the gray‑testing stage
• Reddit: Discussion on DeepSeek’s new 1‑million context window — first reactions from overseas developer communities