DeepSeek V4: Million-Token Context That Actually Works

Most long-context models are benchmarks in search of a use case. DeepSeek V4 flips the script—it delivers 1 million tokens not as a spec-sheet checkbox, but as an operational reality you can actually deploy.

The breakthrough is not just the context length. It is how they got there without torching your inference budget.

The KV Cache Problem Nobody Talks About

Everyone wants to brag about context windows. Few mention that a naive 1M token implementation would need 83.9 GiB of KV cache per sequence using standard attention. That is not a deployment. That is a denial-of-service attack on your GPU memory.

DeepSeek's fix is a hybrid attention architecture that compresses the KV cache by nearly 9x. They use shared key-value vectors across layers, compressed KV streams, and sparse attention on compressed tokens. The sliding window for nearby context stays at 128 tokens—enough for local coherence without the memory bomb.

The numbers: at 1M tokens, V4 needs 9.62 GiB instead of 83.9. With FP4 index cache and FP8 attention, you get another ~2x reduction. That is the difference between works on an 8xH100 node and works on a single node period.

Two Models, One Architecture

V4 ships as Pro and Flash variants. The Pro runs 1.6T total parameters with 49B active per token. Flash scales down to 284B/13B. Both use the same attention architecture, both hit 1M context, both drop to 10 percent of the KV memory you would expect.

The pricing tells the story: Pro at $1.74/$3.48 per million tokens, Flash at $0.14/$0.28. That is not just aggressive pricing—it is a bet that long-context inference can be commoditized if you solve the memory bottleneck.

Why This Matters for Agents

Agent workflows are the stress test for context. A coding agent holding a 300K line codebase in context. A research agent tracking citations across 50 papers. A customer service agent with a year of interaction history.

These are not niche use cases. They are the core value proposition of agentic systems, and they have been blocked by context limitations that forced constant retrieval, reranking, and state fragmentation.

DeepSeek V4's compressed attention means you can keep state resident. No round-trips to vector databases mid-turn. No approximating what matters. Just the full context, available at inference time.

Independent benchmarks show V4 Pro leading open-weight models on agentic tasks—ahead of Kimi K2.6, GLM-5.1, and MiniMax-M2.7 on the GDPval-AA workbench. The Flash variant holds its own at 12x lower cost.

The MoE Efficiency Play

V4's 1.6T parameter count is misleading. With 49B active per token, you are not paying the full FLOPs bill. The key is the routing: DeepSeek uses learned hash routing derived from 2021 ParlAI work, refined through their MoE iterations since V2.

The result is inference throughput that does not collapse under load. Day-zero vLLM integration, MLX quants for Apple Silicon, and a checkpoint that fits on 8xB200s in mixed FP4/FP8. This is a model designed for production, not leaderboard farming.

The Geopolitical Shadow

There is a reason DeepSeek released both base and instruct versions under MIT license, with day-one support for Huawei Ascend chips. The technical achievement exists in a context of compute sovereignty—proving that state-of-the-art long-context models do not require NVIDIA-locked stacks.

Whether that independence narrative holds depends on whether Ascend 950 supernodes can actually scale this year. DeepSeek's own pricing hints at a 50 percent-plus drop once they do. But the architecture is portable enough that it is already running on Blackwell, MI355, and consumer Macs via quantization.

What I Would Actually Use This For

If you are building agents that need persistent, deep context—code review across repositories, legal analysis across case files, research synthesis across thousands of papers—V4 is the first open-weight model where the context window is not the bottleneck.

The hallucination rate on long-context retrieval is still there—about 94 percent on Omniscience benchmarks. The token burn on reasoning tasks is real—V4 Pro uses 190M output tokens on the AA Index versus Flash's 240M. But you can now test these tradeoffs empirically, because the infrastructure actually supports the experiments.

That is the shift. Long context moved from coming soon to ship it. The rest is just optimization.