DeepSeek-V3.1
#192 v Frontier jazykové modelydeepseek · v3.1 · od 2025-08-21 · 11× · naposledy 30. 6. 2026
3
Momentum
DeepSeek-V3.1 is a hybrid frontier language model by DeepSeek with 671 billion total parameters (37B active per token, MoE architecture), released in August 2025. It is DeepSeek's first model to unify "thinking" and "non-thinking" modes in a single checkpoint, combining chat and reasoning model capabilities. The model was built on an extended long-context training process up to 128K tokens and optimized for tool use and agentic workflows. It is available as an open-weight model under the MIT license and is also accessible via the DeepSeek API and third-party providers.
Vývoj momenta
04.04.03.07.
Vlastnosti
| Context Window (Tokens) | 128K tokens (native context via two-phase long-context training: 630B tokens for the 32K phase, 209B tokens for the 128K phase); available as 163,840 tokens on OpenRouter |
| License | MIT License (open-weight, commercial use permitted, self-hosting possible) |
| Multimodality | Text only (no native image, audio, or video input); supports structured JSON output, function/tool calling, code and search agents |
| Price | Free to use via chat.deepseek.com (web app); API: pay-per-token from $0.21/1M input tokens (depending on provider); open-weight download free (self-hosting) |
| Price per 1M Tokens | From $0.21 / 1M input tokens, $0.79 / 1M output tokens (via OpenRouter/DeepInfra); DeepSeek direct (as of March 2026): $0.27 input / $1.10 output per 1M tokens |
| Release Date | August 21, 2025 |