DeepSeek's Million-Token Context Explained — 1M Input · 384K Output

The full V4 lineup offers 1M token context with a max output of 384K; affordable long context comes from the architecture, not subsidies.

Token input

Approx. 750,000 Chinese characters / an entire mid-sized code repository

384K

Max output

Generate full projects and long reports in a single pass

27%

FLOPs (vs V3.2)

Relative compute overhead under 1M load

Hybrid attention: CSA + HCA

Compressed Sparse Attention (CSA)

Compresses attention computation onto sparse key positions, so for long sequences compute grows far more slowly with length than with standard attention—this is what makes the 1M window feasible.

Heavily Compressed Attention (HCA)

Applies heavily compressed memory to "distant" context, reducing the KV cache to roughly 10% of V3.2's; paired with a 128K precise window to preserve recall quality.

Real-World Workload Recommendations

Scenario	Recommended Tier	Rationale
Whole-Repo Coding Agent	deepseek-flash	1M fits the entire repo + vision reads architecture diagrams; 2500 concurrent requests suit CI batch processing
RAG over hundreds of pages of documents	deepseek-flash	Cache hits at $0.003, bringing the cost of repeated retrieval down to nearly zero
Ultra-long offline reasoning	deepseek-v4-pro	The 1.6T flagship handles the hardest cross-file reasoning, and batch processing isn't subject to concurrency limits

Limitation note: Both official and third-party evaluations point out that 1M retrieval recall still lags behind Claude Opus 4.6 (MRCR); its breadth of world knowledge is weaker than Gemini-3.1-Pro—for factual Q&A, pairing it with retrieval is recommended.