Long-horizon reasoning and agentic automation stress both memory (KV cache) and per-token inference cost; this release targets that gap by combining a sparse/hybrid attention design, MoE scaling, and speculative decoding so production services can handle million-token contexts with far lower runtime cost.
Key Capabilities
- Million-token context support and compact KV usage: MoE design activates ~49B parameters per token while the hybrid attention reduces single-token FLOPs and KV cache footprint compared with prior DeepSeek versions, enabling practical 1M-token inference workloads.
- Speculative decoding for lower latency: DSpark speculative decoding is integrated and exposed via vLLM and SGLang flags to reduce wall-clock latency during autoregressive generation in production deployments.
- Production-focused precision and runtime options: Checkpoints and recipes support FP4 + FP8 mixed precision for experts and FP8 for other weights, with vLLM/SGLang deployment recipes and recommended sampling settings for agentic scenarios.
- Agent and coding-agent competency: Tuned evaluation and configuration for agentic tasks (reasoning effort levels, tooling integrations) and shows strong scores on internal and public agent benchmarks compared to earlier preview releases.
Who it's for and tradeoffs
Great fit if you run long-context or agentic services that need practical million-token reasoning (e.g., code agents, multi-document assistants, long-dialogue systems) and can provision MoE-capable inference stacks (vLLM, SGLang, or vendor runtimes with MoE support). Look elsewhere if you need a tiny on-device model, guaranteed open-source community checkpoints without proprietary runtime dependencies, or if your deployment cannot support MoE/expert-parallel execution and speculative decoding engineering.
Where it fits
Positioned as a production-oriented, research-informed LLM: it trades very large total parameter count (1.6T) for a smaller activated-per-token footprint and specialized attention + MoE mechanisms, aiming to close the gap between academic long-context prototypes and deployable agentic systems. Expect nontrivial infra complexity (expert parallelism, mixed-precision tooling) in exchange for substantially larger effective context windows and improved agent benchmarks.
