DeepSeek V3.1

Acerca de DeepSeek

Fundado en 2023 · Hangzhou, China

DeepSeek AI was founded in July 2023 in Hangzhou by Liang Wenfeng, also co-founder of the High-Flyer quantitative hedge fund. The fund's pre-export-control GPU cluster financed DeepSeek's training runs. The lab is known for transparent technical reports and an aggressive open-weights strategy under MIT license. Releases include DeepSeek Coder (Nov 2023), DeepSeek LLM 67B (Jan 2024), DeepSeekMath with GRPO (Feb 2024), DeepSeek V2 introducing Multi-head Latent Attention (May 2024), DeepSeek V3 in December 2024 trained for ~$5.6M of GPU-hours, DeepSeek R1 in January 2025 and DeepSeek V3.1 in 2025 as an incremental update consolidating the base model and the R1 reasoning capabilities into a unified hybrid model. The company has roughly 200 researchers and is privately backed by High-Flyer rather than venture capital. Its V3/R1 release triggered a global re-evaluation of frontier-AI training economics and a notable stock-market move in late January 2025.

Arquitectura

Sparse Mixture-of-Experts Transformer (hybrid base + thinking modes)

DeepSeek V3.1 is a 2025 update of the V3 base that unifies chat (non-thinking) and reasoning (thinking) modes into a single hybrid checkpoint. It retains the V3 architecture - a Sparse MoE Transformer with 671B total and 37B active parameters using DeepSeekMoE routing and Multi-head Latent Attention - but expands the pretraining corpus and updates the post-training recipe. According to DeepSeek's release notes, V3.1 was continually pretrained on ~840B additional tokens of long-context data, extending effective context handling and improving long-document recall within the 128K window. Post-training merged the V3 chat data with R1-style long-CoT reasoning data plus tool-use and agentic trajectories. V3.1 exposes two operating modes selected via the chat template: 'non-thinking' (V3-style fast responses) and 'thinking' (R1-style chain-of-thought before the answer), letting developers choose per request. Tool use and function calling are first-class and improved over both V3 and R1. The model also includes targeted strengthening on coding, agent benchmarks (SWE-bench, Terminal-Bench), and search-augmented reasoning. Weights are released under MIT license and the official DeepSeek API hosts both V3.1 and V3.1-Terminus checkpoints.

Parámetros

671B total, 37B active per token (extended for V3.1)

Contexto

128.000 tokens

Capacidades

• Hybrid model: switchable thinking / non-thinking modes in one checkpoint
• 671B-parameter MoE with 37B active per token
• 128K context window, retrained on ~840B additional long-context tokens
• Strong agentic and tool-use performance on SWE-bench Verified and Terminal-Bench
• Function calling and parallel tool calls
• Long-CoT reasoning inherited from R1
• Open weights under MIT license
• DeepSeek API approximately 1/20th the cost of GPT-4o-class models
• Compatible with vLLM, SGLang, llama.cpp, HuggingFace
• Improved code editing and diff-format generation
• Best for: budget-conscious agentic workloads, coding, hybrid reasoning, on-prem enterprise.

Entrenamiento y licencia

Built on V3's 14.8T-token base, then continually pretrained on roughly 840B additional tokens biased toward long-context documents and code. Post-training combines V3 chat data with R1-style long-CoT and agentic tool-use trajectories.

Licencia: MIT license for weights, code and tokenizer; commercial use permitted.

Pruebas de seguridad: Limited published safety evaluations. As with V3 and R1, politically sensitive topics aligned to Chinese regulations are filtered while general-purpose refusal rates remain low.