Most production agent workflows struggle with balancing long context, tool calls and low-latency decoding. This release targets that intersection: a model checkpoint configured for speculative decoding (DSpark), multi-hundred‑thousand‑token contexts, and OpenAI Responses/Codex compatibility, delivering notably higher agentic and code-agent benchmark scores without simply increasing activated parameter counts.
Key Capabilities
- Agent-focused decoding: ships with a speculative decoding module (DSpark) and provides vLLM and SGLang recipes to enable low-latency speculative generation in agent loops. This reduces latency for multi-step tool-invoking interactions.
- Long-context support: recommended sampling and configuration for
high/maxreasoning effort levels and guidance for outputs up to 384K tokens, making it practical for very long dialogues, documents or multi-step agent traces. - Strong code-agent performance: reports higher scores than earlier preview checkpoints across public and internal agent/coding benchmarks (e.g., Terminal Bench 82.7, Cybergym 76.7, DSBench-FullStack 68.7), indicating improved multi-step reasoning and tool orchestration.
- OpenAI compatibility and integration: includes OpenAI Responses API–format encoding helpers, Codex compatibility notes, and example encoders/parsers to ease integration into existing agent stacks.
Who it's for & trade-offs
Great fit if you operate or prototype agentic systems that need long-context state, tool calling, and lower end-to-end latency (e.g., code synthesis agents, multi-step automation agents, long-document assistants). The model card and repo provide deployment recipes for vLLM/SGLang and suggestions for sampling (temperature=1.0, top_p=0.95 for agent scenarios). Look elsewhere if you need a tiny on-device model or a drop-in smaller-weight checkpoint for CPU inference with no specialized serving infra—achieving the published latencies and multi-token context benefits requires nontrivial inference infrastructure (vLLM/SGLang, MoE/expert backends, GPU memory tuning). Note some benchmark sets cited are internal to DeepSeek; replicate results on your workload before production roll-out.
Additional notes: the release is distributed under the MIT license, includes encoding utilities to map OpenAI-style messages to model prompts, and documents recommended reasoning_effort modes (low, high, max).
