DeepSeek logo

DeepSeek V3.1 Terminus

A 671B MoE refinement of V3.1 (37B active) with improved language consistency and agentic tool use, under the MIT license.

Params
671B, 37B active
Context
131K tokens
Max output
8K
Released
License
MIT
Visit DeepSeek Compare vs other LLMs →

Estimated cost to self-host

DeepSeek V3.1 Terminus needs about 373 GB of GPU memory at 4-bit with 32K context. The cheapest rental that fits is 8× A100 at about $10,714 a month.

Single node ≤ 8
What it costs to self-host DeepSeek V3.1 Terminus at each precision: the memory it needs, the cheapest rental that holds it, that rental's month and the API volume it pays off at
Precision Memory Cheapest fit (32K context) Cost /mo
4-bit INT4 / FP4 373 GB Nvidia logo 8× A100 $10,714
8-bit FP8 / INT8 674 GB Nvidia logo 8× RTX PRO 6000 $13,709
16-bit FP16 / BF16 1,346 GB Nvidia logo 8× B300 $54,720

Estimates based on median on-demand rates for Nvidia GPUs. Memory is weights plus KV cache for one request, using FP8 KV cache where supported and FP16/BF16 otherwise. No guarantee of runtime support, usable performance, or that a matching quantized build exists. Pricing methodology.

Capabilities

What DeepSeek V3.1 Terminus accepts and can do, as published by DeepSeek.

  • Text Accepts and generates natural-language text.
  • Reasoning Thinks before it answers, always on or as a switchable mode.
  • Function calling Connect to external tools, APIs, and systems.
  • Structured output Return responses in structured formats like JSON.

Common questions

What is DeepSeek V3.1 Terminus good for?

The same thinking and non-thinking agent work as V3.1, with what DeepSeek says is less mixed Chinese and English output and better code and search agents. MIT-licensed weights, on a multi-GPU node.

When is DeepSeek V3.1 Terminus not a good fit?

Image input, and long single responses, since its maximum output is short. Running the weights yourself takes a multi-GPU node.

What is the cheapest way to run DeepSeek V3.1 Terminus?

No provider we track hosts it; self-hosting on the cheapest rental that fits (8x A100, $14.88 per hour) costs about $10,714 a month.

How does DeepSeek V3.1 Terminus compare with DeepSeek V3.1?

DeepSeek calls DeepSeek V3.1 Terminus a fix release for DeepSeek V3.1: it cuts mixed Chinese and English text and stray characters in output and tunes the code and search agents further. Same 131K-token context and text input. DeepSeek V3.1 Terminus came out a month later.

Can I self-host DeepSeek V3.1 Terminus?

Yes. The weights are MIT licensed. At 4-bit it needs about 373 GB of GPU memory, which starts at roughly $10,714 a month on the cheapest rental that fits.

More from DeepSeek

All DeepSeek models