DeepSeek released V3.1-Terminus in August 2026, the company’s latest hybrid AI model combining reasoning and speed in a single system. The model cuts hallucinations by 38% and improves multi-step reasoning by 43% over earlier versions.

The new release moves DeepSeek toward hardware independence. The model runs on domestic Chinese accelerators instead of relying on NVIDIA chips alone, a significant shift for a startup competing globally on technology.

Dual-Mode Performance: Think and Fast

V3.1-Terminus supports two modes. The Think mode activates deep reasoning chains for complex problems. Fast mode skips reasoning for quick answers. Users pick the trade-off they need for each query.

This flexibility matters because reasoning models burn inference tokens. Maths problems and coding tasks benefit from thinking. Factual lookup and summarization don’t. V3.1 lets users control when to pay the cost.

Context Window and Parameter Efficiency

The model handles 128,000-token context windows, matching competitors like OpenAI’s GPT-4 and Google’s Gemini. It uses only 37 billion active parameters out of 671 billion total through mixture-of-experts architecture, making it efficient to run.

Smaller active parameters mean lower latency and cheaper inference. DeepSeek built V3.1 to cost less to operate than rivals while maintaining competitive benchmark scores.

Free Access and Benchmark Results

DeepSeek makes V3.1 free through its API. The model outperforms DeepSeek-R1 and earlier V3 versions across benchmarks. It trails only Google’s closed Gemini 3.1-Pro on knowledge tasks but beats it on maths and coding.

For open-source models, V3.1 sets a new standard. Researchers and startups can access frontier reasoning capability without licensing fees.

V3.1 signals DeepSeek’s arrival as a genuine competitor, not just a fast follower.

FYI (keeping you in the loop)

Is DeepSeek really open-source?

The model weights are free and open, but the training and fine-tuning methods aren’t fully disclosed. It’s more open than closed models like GPT-5, less open than purely academic models.