DeepSeek-V4登場、エージェントが実用可能な100万トークン文脈を実現 DeepSeek-V4: a million-token context that agents can actually use

DeepSeek, the Chinese lab behind a fast-growing family of open-weight large language models, has unveiled DeepSeek-V4, its newest flagship release. The headline feature is a context window pushing into the million-token range, paired with a stated focus on making that long context actually useful for agentic workloads rather than serving as a benchmark trophy.

Long-context claims have become almost routine in the LLM space, but practitioners have repeatedly run into the same failure modes. Models tend to drop information buried in the middle of a prompt — the well-documented "lost in the middle" problem — and inference cost and latency balloon as token counts climb. According to DeepSeek's accompanying notes, V4 reworks the attention mechanism and key-value cache handling to improve retrieval and reference accuracy deep inside long inputs, with the explicit goal of supporting multi-step agent loops that involve tool calls, intermediate reasoning traces, and accumulated state.

The practical motivation is clear. Demand for genuinely usable long context has grown across several workloads: whole-repository code understanding and refactoring, research agents that traverse many documents, customer-facing assistants that need durable memory across sessions, and pipelines that ingest large logs or transcripts in a single pass. In each of these, the bottleneck has rarely been raw window size — it has been whether the model can reliably pick out the right span when the window is mostly full.

On the architecture side, DeepSeek-V4 appears to continue the Mixture-of-Experts direction the lab has pursued since V2 and V3, with sparse expert activation keeping per-token compute manageable even as parameter counts grow. Combined with the long-context optimizations, this positions V4 as an attempt to push the MoE recipe further into territory previously dominated by closed frontier models. Specific numbers around active parameters, expert routing, and benchmark scores will need independent verification as the community runs its own evaluations.

The competitive backdrop matters here. Google's Gemini 1.5 Pro put one-to-two-million-token contexts into production use, and both Anthropic's Claude family and OpenAI's GPT line have steadily extended their windows. Open-weight options at this scale, however, have been comparatively scarce. Alibaba's Qwen series has shipped long-context variants, and hybrid approaches incorporating state-space models such as Mamba continue to attract research interest as a way to escape the quadratic cost of standard attention. DeepSeek-V4 slots into that landscape as an MoE-first answer that prioritizes long-range fidelity over architectural novelty.

The choice to publish on Hugging Face carries real weight for developers who want to run or fine-tune the model themselves. Inference stacks such as vLLM and SGLang have been actively adding support for MoE routing and long-context KV cache optimizations, including paged attention, prefix caching, and disaggregated prefill. That trajectory has steadily lowered the barrier to operating million-token-class models outside hyperscaler APIs, though the engineering remains nontrivial.

In practice, memory footprint and throughput will likely be the limiting factors for most teams. Holding a million-token KV cache in GPU memory is expensive even with grouped-query or multi-head latent attention schemes, and prefill latency on prompts of that length can be measured in tens of seconds without careful batching. Quantization down to 8- or 4-bit weights, along with context-compression techniques such as summarization, retrieval pre-filtering, or learned token pruning, may turn out to be necessary companions rather than optional tweaks for cost-sensitive deployments.

There are also open questions the release itself does not fully resolve. How V4 behaves on adversarial long-context evaluations such as needle-in-a-haystack variants with distractors, RULER, or agent-oriented benchmarks like SWE-bench and τ-bench will determine whether the agent-readiness pitch holds up. Licensing terms, training data disclosures, and tool-use fine-tunes will similarly shape adoption in regulated environments. For now, DeepSeek-V4 reads as a credible attempt to close the gap between nominal context length and the long-horizon reasoning that real agent systems demand — and, importantly, to do so with weights that anyone can download.