DeepSeek V3.1 Terminus: BEST Opensource Agentic Model! Powerful, Fast, & Cheap! (Fully Tested)

Key Concepts:

• Deepseek V3.1 Terminus: An updated open-source language model by Deepseek.
• Code Agent: The model's capability to generate and understand code.
• Search Agent: The model's ability to use search engines to gather information.
• Reasoning Mode: A mode where the model thinks more extensively before responding.
• Agentic Performance: The model's ability to perform tasks autonomously using tools.
• Context Window: The amount of text the model can consider at once (131k in this case).
• Token Pricing: The cost per million input and output tokens (27 cents and $1 respectively).
• Benchmarks: Standardized tests used to evaluate the model's performance (e.g., Sway Verified, MMLU, Live Codebench, Code Force, ADR Polygon).
• Kilo Code: A platform offering free credits to access coding agents like Deepseek Terminus.
• SVG Code: A vector image format used for creating graphics.

1. Introduction of Deepseek V3.1 Terminus

• Deepseek team released Deepseek V3.1 Terminus model, an update to their previous model.
• The new release addresses user-reported issues and enhances existing capabilities.
• Key improvements include better code and search agent performance and improved language consistency.
• The Terminus model provides more stable and reliable output across various benchmarks.

2. Performance Improvements and Trade-offs

• The Terminus model shows noticeable improvements compared to the earlier Deepseek V3.1, but it's not a groundbreaking upgrade.
• Benchmarks show a slight increase in Sway Verified (agentic tool use), MMLU humanities last exam (reasoning), and Live Codebench.
• Performance gains are particularly noticeable with reasoning mode enabled without tool use.
• There's a slight decrease in performance on certain benchmarks like Code Force and ADR Polygon due to optimization trade-offs.
• The team may have focused on improving stability and reasoning capabilities, leading to these decreases.
• Fine-tuning for agentic tool use benchmarks might explain the increase in Sway Verified and Terminal Bench.

3. Cost Efficiency and Context Window

• Deepseek V3.1 Terminus is a cost-efficient open-source model.
• Pricing: 27 cents per 1 million input tokens and $1 per 1 million output tokens.
• Context window: 131k, with a 65.6K max output.
• Despite the context window not being massive compared to other models, it delivers strong agentic performance and reasoning.

4. HubSpot's "Master AI Agents in 2025" Guide

• HubSpot offers a free guide called "Master AI Agents in 2025: The Strategic Advantage."
• The guide helps users understand where to start with AI agents and which applications deliver real value.
• It covers automating marketing workflows, accelerating sales, and streamlining operations.
• It provides a step-by-step blueprint for selecting, deploying, and measuring AI agents.

5. Accessing and Using Deepseek V3.1 Terminus

• The model can be accessed via Deepseek's chatbot, which has the Terminus model implemented.
• It can also be accessed via API provided by Deepseek or through external providers like Open Router.

6. Benchmark Tests and Results

• SAS Landing Page Creation: The model was tasked with creating a SAS landing page with many features. The result was a typical AI-generated landing page with expected components, a well-structured layout, and decent appearance. It performed better than the previous version.
• Modern Web Browser Creation (using Kilo Code): The model was asked to create a modern web browser similar to Chrome. It generated a browser called "Nexus" with a main dashboard, extension store, and settings tab. The base structure was created, but the functionality was limited.
• Portfolio Management Proposal: The model was asked to draft a portfolio management proposal for a truck driver making $65K a year aiming to retire in 30 years. The chatbot's answer was less detailed compared to the Open Router provider. The Open Router response provided a more structured plan, considering inflation and tailoring the plan to the specific income and timeline.
• Butterfly Creation in SVG Code: The model was tasked with generating a symmetrical butterfly in SVG code. The initial attempts failed to create a recognizable butterfly. After multiple attempts, the best result was an animated attempt that didn't create the main structure of the wings.
• Minecraft Clone Creation: The model was asked to create a Minecraft clone. It generated a 3D clone with sound, block placement, and breaking capabilities. However, the functions were not fully functional, with the user falling out of the map.

7. Conclusion

• Deepseek V3.1 Terminus is a cost-efficient and high-performing model with improvements in coding, reasoning, and search agent capabilities.
• While there are some performance trade-offs, it offers meaningful upgrades in key areas.
• The model can be easily accessed through Deepseek's chatbot or API.
• The benchmark tests showed mixed results, with successes in SAS landing page creation, browser creation, and portfolio management, but failures in butterfly creation. The Minecraft clone was partially successful.
• The model is recommended for users looking for an open-source model with strong agentic performance and reasoning.