For most of 2025, "long context" meant 128,000 tokens if you were lucky and a steep bill if you wanted more. DeepSeek's V4, released in April 2026, made 1 million tokens the default and gave the weights away.
The move continues a pattern: a Chinese lab setting the open-source baseline on context length and price, then letting developers decide whether to self-host or call an API.
What V4 actually is
DeepSeek V4 comes in two sizes:
- DeepSeek-V4-Pro — 1.6 trillion total parameters, 49 billion active per token.
- DeepSeek-V4-Flash — 284 billion total parameters, 13 billion active per token.
Both use a Mixture-of-Experts (MoE) architecture that keeps the V3 lineage (DeepSeekMoE plus multi-token prediction) and adds new attention and optimization work: a hybrid attention design combining compressed and heavily compressed attention, DeepSeek Sparse Attention (DSA), manifold-constrained hyper-connections (mHC), and the Muon optimizer.
The headline number is context: 1 million tokens is now the standard window across DeepSeek's official services.
Open weights, MIT license
DeepSeek released V4 as open weights under an MIT license, distributed through its open-source repository and an API. The API keeps the same base URL as before; developers switch by changing the model name to deepseek-v4-pro or deepseek-v4-flash. It supports both OpenAI-style ChatCompletions and Anthropic-style API calls.
The older deepseek-chat and deepseek-reasoner endpoints were retired after July 24, 2026 (15:59 UTC), with traffic routed to V4-Flash in thinking and non-thinking modes.
Where DeepSeek says it lands
Per DeepSeek's own benchmarking, V4-Pro leads open models on world knowledge (trailing only Gemini-3.1-Pro among closed models), tops open-source results on agentic coding, and beats all current open models on math, STEM and competition coding. V4-Flash is positioned as close on reasoning while cheaper and faster. The models integrate with agent tools including Claude Code, OpenClaw and OpenCode.
Why a two-tier line
DeepSeek's split is a pricing strategy dressed as engineering. The Pro tier is the showcase — large, capable, expensive to serve. The Flash tier is the volume product: small enough to run cheaply, fast enough for interactive use, and good enough on simple agent tasks that most callers never need Pro. By giving both away as weights, DeepSeek pressures closed competitors on the exact axis enterprises care about — total cost of ownership — while keeping its own API as the convenient default.
The open-weight angle
For Chinese developers locked out of some US models, an MIT-licensed 1M-context model is more than a benchmark. It is a foundation they can fork, fine-tune and embed without per-call fees or geopolitical risk. That is why V4 matters as corporate strategy, not just model quality: DeepSeek is building a default open stack the way Android built a default mobile stack.
Who feels the pressure
The 1M window lands hardest on retrieval-heavy workflows. Enterprise search, legal document review and codebases with millions of lines are exactly the cases where stitching together a vector store used to be mandatory. DeepSeek is effectively arguing that context length has become cheap enough to replace plumbing with raw window size. That pressures not just closed model vendors but also a whole layer of retrieval and orchestration startups whose main job was managing context that models could not hold.
The catch on cost
Long context is not free. Feeding a million tokens per request multiplies the tokens billed and the memory held per session, so a single agent loop that re-sends the window on every step can get expensive fast. DeepSeek's sparse-attention work is meant to blunt this, but the economic break-even still depends on how much of that window a task actually uses.
Deployment takeaways
According to Li An, chief scientist at BrainNet (脑机网), a 1M-token default changes how teams build agents: instead of juggling retrieval pipelines, they can drop entire codebases or document sets into one context, which trades latency and cost for dramatically simpler architecture.
Honest limitations
The performance claims above are DeepSeek's own and should be read as marketing-grade until reproduced by independent labs. "1 million tokens" is a context ceiling, not a promise of stable quality at that length; long-context accuracy often degrades deep into the window. The MIT weights let you self-host, but running a 1.6T-parameter model still demands serious GPU memory and engineering. Dates and parameter counts here come from DeepSeek's official model card and release page; we found no contradictory reporting, but the figures are single-source (the lab itself).
What readers can do now
- If you run agents, test V4-Flash on a long-codebase task where retrieval used to fail, and compare cost against your current model.
- For privacy-sensitive work, self-host the MIT weights rather than sending data through the API.
- Set a calendar check for independent benchmarks in the next quarter before committing production traffic to V4-Pro.