DeepSeek V4 Preview and the 1M-context open model race
DeepSeek's V4 Preview is another signal that Chinese AI labs are competing hard on context length, active parameters, and cost.
Having a million tokens of context doesn’t help if the model can’t reliably use what’s inside it. Long context is becoming table stakes — the real differentiator is what happens at that length.
Source
By: DeepSeek
Published: Apr 24 2026
Type: Release announcement
DeepSeek matters because it changed how developers think about open models and cost-performance. V4 Preview continues the pattern: large context, open weights, and two variants split along the cost-capability axis.
How V4 splits the open-model trade-off into two paths
DeepSeek-V4 ships as two models. V4-Pro has 1.6 trillion total parameters with 49 billion active, using sparse activation (mixture-of-experts) to keep the total servable. It targets world-class reasoning — the release claims open-source state-of-the-art on agentic coding benchmarks, leading all open models in math/STEM/coding, and rivalling top closed-source models in rich world knowledge (trailing only Gemini 3.1 Pro on that axis).
V4-Flash has 284 billion total parameters with 13 billion active. It’s the cost-sensitive path: reasoning capabilities that closely approach V4-Pro, on-par performance with V4-Pro on simple agent tasks, smaller parameter size, faster response times, and cheaper API pricing. The two models together give developers the same choice pattern that’s now standard — a strong model for hard tasks and a fast/cheap model for everything else — but with 1M context as the default on both.
The structural innovation is in how that context length works. V4 introduces token-wise compression and DeepSeek Sparse Attention (DSA), which the release calls “world-leading” for reducing compute and memory costs at long context. This matters because 1M context on paper means nothing if every token in that window costs the same compute as the first — the whole point of sparse attention is that it doesn’t.
Both models support thinking and non-thinking modes, OpenAI ChatCompletions and Anthropic API compatibility, and integrate with coding agents (Claude Code, OpenClaw, OpenCode are named). DeepSeek also mentions using V4 for in-house agentic coding, which is a concrete signal that the model is being tested in real developer workflows, not just benchmarks.
The practical question is whether 1M context with sparse attention actually retrieves accurately across the full window. Long context has a known failure mode: the model accepts the tokens but doesn’t reliably use information in the middle. DeepSeek’s sparse attention claims to solve the compute and memory side. Whether it also solves the retrieval accuracy side — especially during multi-turn agent loops where context accumulates across steps — is what actual usage will reveal. Latency at 1M context, tool use reliability across long sessions, and cost at sustained agent volumes are the metrics that matter more than the context-length headline.
What I want to remember
Remember: Long context is becoming normal, but it’s not enough by itself. Retrieval quality inside that window, tool use stability, latency at full length, and cost per token across a full agent loop decide whether a 1M-context model is actually usable in production.