DeepSeek V3.1

2025-08-21 · Publisher : DeepSeek · Domain : Artificial Intelligence

DeepSeek releases V3.1, a 671-billion-parameter model that combines a thinking mode and a non-thinking mode in a single set of weights, with clear advances for coding and search agents.

DeepSeek releases V3.1, a 671-billion-parameter model that combines a thinking mode and a non-thinking mode in a single set of weights, with clear advances for coding and search agents. ### Context Until then, DeepSeek maintained two separate lines: V3 for the general-purpose model and R1 for reasoning. V3.1 merges them. The aim is to spare users from choosing between two models and to prepare for what the vendor calls "the agent era". R1-0528 and V3-0324 serve as internal references to measure progress. ### What the model brings - Hybrid inference: the same architecture answers in thinking or non-thinking mode depending on the chat template, and the DeepThink button switches between the two. - Faster thinking: V3.1 in thinking mode reaches its answers in less time than R1-0528, according to the blog post. - Better agent skills: SWE-bench Verified in agent mode at 66.0 (versus 45.4 for V3-0324) and Terminal-bench at 31.3 (versus 13.3). - A long-context extension on the V3 base: 630 billion tokens for the 32K phase (multiplied by 10) and 209 billion for the 128K phase (multiplied by 3.3), for a total of 840 billion tokens. - The UE8M0 FP8 numerical format for weights and activations, presented by the vendor as compatible with the "microscaling" format (details in DeepGEMM). - Support for the Anthropic-format API and strict function calling in beta. ### Architecture and specifications - Type: Mixture-of-Experts, 671 billion parameters in total, 37 billion activated per token. - Context: 128K tokens for both modes. - Modalities: text. Number of experts, base data and knowledge cutoff: not disclosed in the sources consulted. - License: MIT (code and weights), commercial use allowed. - Weights: huggingface.co/deepseek-ai/DeepSeek-V3.1 and DeepSeek-V3.1-Base. - API identifiers: `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking). ### Cost and access | API (USD per million tokens) | Price from September 5, 2025, 4 p.m. UTC | |---|---| | Input, cache hit | 0.07 | | Input, cache miss | 0.56 | | Output | 1.68 | The blog post states that the new pricing grid replaces the previous prices and off-peak discounts on that date; the previous pricing is not quantified there. Self-hosting hardware: not disclosed. ### Benchmarks | Benchmark | V3.1 non-thinking | V3-0324 | V3.1 thinking | R1-0528 | |---|---|---|---|---| | MMLU-Pro (EM) | 83.7 | 81.2 | 84.8 | 85.0 | | GPQA Diamond (Pass@1) | 74.9 | 68.4 | 80.1 | 81.0 | | LiveCodeBench (Pass@1) | 56.4 | 43.0 | 74.8 | 73.3 | | SWE Verified (agent mode) | 66.0 | 45.4 | not published | 44.6 | | Terminal-bench | 31.3 | 13.3 | not published | 5.7 | | AIME 2024 (Pass@1) | 66.3 | 59.4 | 93.1 | 91.4 | | BrowseComp | not published | not published | 30.0 | 8.9 | Source: DeepSeek-V3.1 model card on Hugging Face, scores published by the vendor. This table only compares DeepSeek models. ### Versus closed models The primary source contains no comparison table with a closed model for V3.1: no quantified gap can therefore be established here. The blog post instead positions itself against the vendor's own models: in thinking mode, V3.1 matches or exceeds R1-0528 on LiveCodeBench (74.8 vs 73.3) and AIME 2024 (93.1 vs 91.4), while thinking faster. On price, the September 5, 2025 grid sets output at 1.68 dollars per million tokens. The open choice takes the form here of a single model to deploy for two uses, under the MIT license. ### Key takeaways V3.1 removes the separation between general-purpose model and reasoning model, which simplifies hosting and integration. Its gains on agent tasks make it the first agentic milestone of the DeepSeek line.

Tags : LLM, Open Source, DeepSeek, MIT, MoE, Modèle hybride, Agents