Announcing DeepSeek-V3.1 on GMI Cloud

DeepSeek-V3.1 is the latest upgrade to DeepSeek’s flagship open-weight LLM. The Instruct model is now fully integrated into the GMI Cloud inference engine. It introduces a hybrid inference architecture—supporting both fast, direct responses (“Non-Think” mode) and deep, multi-step reasoning (“Think” mode)—while enabling 128K-token context handling, open-source accessibility, and better integration for tool-using AI agents. 

What’s New in DeepSeek-V3.1

Hybrid Inference: Think & Non-Think Modes

DeepSeek-V3.1 introduces a dual-mode system:

• Non-Thinking Mode → Fast, concise answers for efficiency
  
  
• Thinking Mode → Deep, step-by-step reasoning for complex workflows
  
  

Users can toggle modes via the DeepThink button on the app or web interface.

API & Integration Enhancements

Two Endpoints for Flexibility

• deepseek-chat: optimized for non-thinking responses
  
  
• deepseek-reasoner: built for reasoning-intensive tasks
  
  

Integration Upgrades

• Supports 128K-token context windows for both endpoints
  
  
• Adds Anthropic-style API formatting
  
  
• Enables Strict Function Calling (Beta) for reliable, agent-driven workflows
  
  

Model Architecture Upgrades

Long-Context Pretraining

• Expanded 32K-phase training by 10× to 630B tokens
  
  
• Expanded 128K-phase training by 3.3× to 209B tokens

Efficient Precision Format
Uses UE8M0 FP8 for faster processing speeds and compatibility with micro-scaling formats.

Open-Source Release
Both V3.1 base weights and the full model weights are publicly available on Hugging Face.

Performance Boosts & Agent Capabilities

• Smarter Tool Use → Better multi-step reasoning, API integration, and autonomous workflows
  
  
• Faster “Thinking” Mode → Matches DeepSeek-R1-0528’s accuracy but responds more quickly
  
  
• Improved Agent Behaviors → More reliable search, integration, and orchestration of external tools
  
  

Performance Benchmarks

DeepSeek-V3.1 consistently outperforms earlier versions across code, reasoning, and search benchmarks, showing major gains in SWE-bench, multilingual tasks, and complex search. It also produces longer, higher-quality outputs on reasoning-heavy benchmarks like AIME 2025 and GPQA.

‍

Run DeepSeek-V3.1 on GMI Cloud

You can deploy DeepSeek-V3.1 immediately through our inference engine by following the instructions here.

GMI Cloud provides the infrastructure, tooling, and support needed to deploy DeepSeek-V3.1 at scale. Our inference engine is optimized for large-token throughput and ease of use, enabling rapid integration into production environments

With GMI Cloud, you can:

• Serve DeepSeek-V3.1 via optimized, high-throughput inference backend
• Configure models for batch, streaming, or interactive inference
• Integrate with prompt management, RAG pipelines, and eval tooling
• Connect via simple APIs without additional DevOps effort
• Scale with usage-based pricing and full visibility into performance

At GMI Cloud, we’re excited to offer access to DeepSeek-V3.1 because it delivers open-weight flexibility with cutting-edge reasoning capabilities, empowering developers to build research assistants, knowledge engines, and long-memory AI systems without sacrificing speed or cost efficiency.

​​

Pricing & Availability

DeepSeek-V3.1 is available today via:

• Web app with DeepThink toggle
  
  
• Updated API endpoints
  
  
• GMI Cloud deployment for optimized compute environments
• $0.9/$0.9 with GMI Cloud
  
  

DeepSeek-V3.1 at a Glance

‍

Why It Matters

DeepSeek-V3.1 represents a strategic evolution for AI development:

• Technically, it brings agent-ready inference and long-context handling to open-source models.
  
  
• Politically, its Chinese chip optimization signals an alignment with domestic hardware ecosystems, an important step amid U.S.-China tech tensions.
  
  

Practically, developers gain access to a powerful, flexible model that can toggle between speed and deep reasoning—and now, with GMI Cloud integration, they can scale it effortlessly in production.