DeepSeek's new V3.1 handles thinking and fast chat in one model

DeepSeek-V3.1 is an open-source hybrid model that handles both deep thinking and rapid responses in a single architecture. Clocking in at 671B total parameters with 37B activated per token, it supports a native 128K context window. Weights for both base and chat versions are now freely available on Hugging Face and ModelScope.

What it can do

Instead of running separate models for reasoning and standard conversation, DeepSeek-V3.1 lets you switch modes simply by tweaking the chat template. On the web chat, you can toggle between the two with a single "DeepThink" button.

When running in thinking mode, the model delivers output quality on par with DeepSeek-R1-0528 while cutting down waiting time with faster token generation.

Agent workflows get a major upgrade too. The model posts stronger results across SWE and Terminal-Bench, proving significantly more reliable on multi-step complex search and tool-use tasks.

On the API side, deepseek-chat handles standard queries while deepseek-reasoner manages thinking runs, both maxing out at 128K context. The API now supports the Anthropic API format and introduces Strict Function Calling in beta.

How it works and how to run it

Under the hood, DeepSeek-V3.1 is built on DeepSeek-V3.1-Base using post-training optimizations. The base model underwent continued pretraining with 840B tokens dedicated to long-context extension.

The 32K phase was expanded tenfold to 630B tokens, while the 128K phase grew 3.3 times to 209B tokens. Both weights and activations use the UE8M0 FP8 scale format to ensure hardware microscaling compatibility.

For self-hosting, official recipes are ready for Transformers, vLLM, and SGLang. You can spin up an OpenAI-compatible endpoint with a basic vLLM command or deploy via Docker using SGLang with 32GB shared memory.

Keep in mind that running a 671B MoE model locally demands serious multi-GPU enterprise hardware. If you just want to test it, the web interface is already live, and current API rates with off-peak discounts will hold until September 5, 2025, at 16:00 UTC.

Why it matters

Packaging slow deliberative thinking and quick chat into one model solves a massive operational headache. Developers no longer need to route prompts across different models or pay for redundant GPU allocations.

For agent builders, the tighter tool-calling abilities combined with native Anthropic API compatibility make migrating existing codebases almost frictionless. While local deployment is strictly for high-end server clusters, as an open-weight foundation, V3.1 sets a fierce standard for enterprise-grade open-source AI.