DeepSeek Launches Terminus

DeepSeek, one of the fastest-growing AI companies, has unveiled Terminus, the newest version of its hybrid reasoning model V3.1.

Terminus delivers major improvements in workflow automation, language consistency, large-context handling, and developer affordability.

What Is Terminus?

Terminus builds on DeepSeek’s vision of hybrid reasoning, where AI acts as an intelligent agent rather than just a text generator.

Instead of only producing text, Terminus can use specialized tools such as code execution agents, web search, and structured output functions to complete complex, multi-step tasks with verifiable accuracy.

This marks a clear shift from “sounding smart” to delivering results that can be tested and trusted in real-world applications.

Key Features

Dual Modes: Chat and Reasoner

One of Terminus’s standout features is its dual-mode operation:

Chat Mode – Handles everyday conversations, JSON outputs, and simple function calls. It supports up to 8,000 tokens per output (default 4,000).

Reasoner Mode – Designed for heavy-duty problem solving, with outputs up to 64,000 tokens (default 32,000).

Both modes share a 128,000-token context window, letting users feed hundreds of pages of documents or code into a single session.

When a Reasoner task requires tool use, Terminus automatically shifts the process into Chat mode to ensure accuracy.

Cleaner Language and Stability

Earlier DeepSeek releases sometimes mixed English and Chinese text or introduced random characters.

Terminus fixes this issue with an improved tokenizer and prompt templates.

The result: clean, consistent language output that integrates more smoothly into professional workflows.

Developers note fewer formatting errors and more reliable performance in production environments.

Training and Improvements

Terminus was trained with 840 billion extra tokens, a new tokenizer, and improved scaffolding techniques.

These refinements boost performance across multiple use cases.

However, results show a small drop in Chinese-language browsing, reflecting that this version is more tuned for English-based web tasks.

Deployment and Self-Hosting

Developers can access Terminus via updated demo code on Hugging Face Hub.

Many organizations are already self-hosting for better data control and performance customization.

A small technical issue remains: the self-attention projection parameter does not yet meet the UE8 M0 FP8 scaling format.

DeepSeek says a fix is on the way, but this only affects advanced self-hosted setups, not everyday users.

Benchmark Performance

DeepSeek’s internal tests show strong improvements in tool-using tasks:

• BrowseComp (web search tasks): from 30.0 → 38.5.
• TerminalBench (reasoning workflows): from 31.3 → 36.7.

Other reasoning benchmarks also improved:

• SimpleQA: 93.4 → 96.8.
• SuiVerified: 66.0 → 68.4.
• SWIB Multilingual: 54.5 → 57.8.
• GPQA Diamond: 80.1 → 80.7.
• Humanity’s Last Exam: 15.9 → 21.7.
• However, coding benchmarks dipped slightly: Codeforces fell from 2,091 to 2,046.

DeepSeek admits this is a trade-off, prioritizing agent reliability over raw coding speed.

Open-Source and Pricing

Like earlier DeepSeek models, Terminus is fully open source under the MIT license.

This allows developers and companies to use, modify, and self-host without restrictions something few competitors offer.

Pricing is another disruptive factor:

• $0.07 per million input tokens (cache hits).
• $0.56 per million input tokens (cache misses).
• $168 per million output tokens.
• For comparison, GPT-5 costs about $10 per million output tokens, while Anthropic’s Claude Opus costs around $75 per million.
• Even conservative estimates place DeepSeek at least 10 times cheaper than rivals, making it attractive for startups and large enterprises alike.

Real-World Use Cases

Early testing shows Terminus working well in practical scenarios:

Web Development: Generated SaaS landing page code with structured JSON, animations, and multi-section layouts.

Finance: Handled financial planning prompts with retirement strategies and inflation calculations.

Creative Coding: Built a basic 3D Minecraft-style prototype with block placement, destruction, and sound.

While results weren’t perfect—such as inconsistent SVG drawings—the model shows major progress in structured and creative problem solving.

News Gist

DeepSeek has launched Terminus, an upgraded hybrid AI model with dual Chat/Reasoner modes, improved tool use, cleaner outputs, 128K context, and open-source licensing—delivering affordable, agent-driven workflows for developers, startups, and enterprises worldwide.

FAQs

Q1. What is DeepSeek Terminus?

DeepSeek Terminus is the upgraded V3.1 hybrid reasoning model that integrates tool use, long-context handling, and dual operation modes for smarter AI workflows.

Q2. When was Terminus launched?

Terminus was officially launched on September 22, 2025, as the latest release in DeepSeek’s fast-moving AI model lineup.

Q3. What are its key features?

Key features include Chat and Reasoner modes, a 128K-token context window, improved language consistency, stronger benchmark results, and open-source licensing under MIT.

Q4. How much does it cost?

Pricing starts at $0.07 per million input tokens (cache hit), $0.56 per million input tokens (cache miss), and $168 per million output tokens—much cheaper than GPT-5 or Claude.

Q5. Who can use Terminus?

Developers, startups, and enterprises can access Terminus. It can be self-hosted or used via DeepSeek’s API and Hugging Face demo code.

Q6. Why is Terminus important?

Terminus makes advanced agentic AI systems affordable and open-source, enabling real-world problem solving with lower costs and greater accessibility compared to big competitors.