Today we formally release DeepSeek-V3.1. The upgrade carries three headline changes: a hybrid reasoning architecture — one model supporting both thinking and non-thinking modes; higher thinking efficiency — DeepSeek-V3.1-Think reaches answers in less time than DeepSeek-R1-0528; and stronger agent capability — post-training optimizations deliver substantial gains in tool use and agentic tasks.
The official app and web models have synced to DeepSeek-V3.1, with a “Deep Thinking” button switching freely between thinking and non-thinking modes. The DeepSeek API upgraded in step: deepseek-chat maps to non-thinking mode, deepseek-reasoner to thinking mode, both with context extended to 128K. The API beta also adds strict-mode function calling, guaranteeing outputs conform to the declared function schema. And — the strategically loud part — DeepSeek added Anthropic API format support, letting developers plug DeepSeek-V3.1’s capabilities directly into the Claude Code framework.
Tool Calling and Agent Support, Measured
Coding agents: on SWE (evaluated with an internal framework requiring fewer rounds than the open-source OpenHands setup) and Terminal-Bench (official Terminus 1 framework), DeepSeek-V3.1 improves markedly over previous DeepSeek-series models on both code repair and complex command-line tasks.
Search agents: across multiple search-evaluation metrics, V3.1 advances substantially — on BrowseComp (complex multi-step search) and HLE (multidisciplinary expert-level questions), it now leads R1-0528 by a wide margin, with test calls running through commercial search APIs plus web filtering over a 128K context window.
Thinking Efficiency: Same Scores, Fewer Tokens
After chain-of-thought compression training, V3.1-Think matches R1-0528’s average task performance while emitting 20-50 percent fewer output tokens. The score lines hold nearly level — AIME 2025: 87.5 versus 88.4; GPQA: 81 versus 80.1; LiveCodeBench: 73.3 versus 74.8 — while the token-consumption gap widens in V3.1’s favor. In non-thinking mode too, output length came under control: DeepSeek-V3.1 holds the same performance as DeepSeek-V3-0324 at visibly shorter outputs. Efficiency as a first-class feature, not an afterthought — and for agent workloads, the difference between a usable budget and an unusable one.
Open Weights and API
The V3.1 Base model was re-expanded from V3 with an additional 840 billion tokens of training. Both the Base and post-trained models are open on Hugging Face and ModelScope: DeepSeek-V3.1-Base and DeepSeek-V3.1 on Hugging Face, with mirrors on ModelScope.
The search-agent results deserve a closer look because they mark a threshold. Search agents fail differently from chatbots: they chain a dozen small decisions — which query, which results to trust, when to stop searching and start answering — and every link can silently poison the final answer. DeepSeek benchmarked this the production-hard way: real commercial search APIs, real web filtering, real context pressure. Beating R1-0528 by a wide margin on BrowseComp and HLE under those conditions means the model improved not at answering questions but at conducting investigations — the skill that separates a demo agent from one you would let touch your inbox. Combined with the strict-mode function calling (which guarantees tool arguments parse against the declared schema, eliminating an entire class of runtime failures) and the doubled context window, V3.1 arrives as a complete agent substrate rather than a chat model with agent aspirations.
The Anthropic API compatibility also reads differently depending on where you sit. For developers, it is pure leverage: Claude Code’s harness, permissions, and tool integrations are the most polished agentic runtime available, and being able to point it at an open-weights model converts an API subscription question into a hosting decision. For Anthropic, it is competitive pressure of the most concrete kind — the interface layer it built becoming an open standard that cheaper models can speak. And for the industry, it extends a pattern this publication has tracked all year: Chinese labs competing not by inventing new interfaces but by perfectly implementing the de facto ones. The historical analogy is deliberate — this is how commodity ecosystems have always attacked platform incumbents, by being fluent in their language and aggressive on price. Whether the frontier labs respond by opening their own protocols or by accelerating capability distance will shape the agent market’s next chapter.
Why This Release Matters
Read the changelog as a strategy document and the direction is unmistakable: DeepSeek is repositioning from “the cheap chat model” to “the agent infrastructure.” The hybrid reasoning architecture acknowledges that production agents need both modes — fast talk for conversation, slow thought for planning — in a single deployable model rather than two SKUs. The thinking-efficiency work targets the exact cost line that agentic workloads blow through, since an agent that thinks before every tool call can burn tokens faster than it answers. The strict function calling and 128K context serve the same customer. And the Anthropic API compatibility is the boldest stroke: by speaking Claude’s protocol, DeepSeek makes itself a drop-in alternative inside the most popular agentic coding framework on earth — an open-weights model inviting developers to run the Claude Code experience on their own hardware, at DeepSeek’s prices. For the competitive landscape, V3.1 is the release where the efficiency-first lab formally joined the agent race — not by matching the frontier’s benchmarks alone, but by matching its interface. Weights: huggingface.co/deepseek-ai/DeepSeek-V3.1. Docs: api-docs.deepseek.com. Source: DeepSeek official announcement. For teams evaluating it now, the shortest test path is also the most telling: wire it into Claude Code via the new API compatibility, run your own SWE or search workload, and compare both the answers and the invoice against your current provider — V3.1 is built to win that specific comparison.
