Deepseek

DeepSeek

DeepSeek (from China) disrupted the market in late 2024/2025 by releasing DeepSeek-V3 and R1 (Reasoning) with performance matching Claude/GPT-4 at 1/10th the cost.

When to Use

• Cost Efficiency: The API is incredibly cheap.
• Reasoning: DeepSeek-R1 uses Chain-of-Thought reinforcement learning (like OpenAI o1) but is open weights.
• Coding: DeepSeek-Coder-V2 is a top-tier coding model.

Core Concepts

MLA (Multi-Head Latent Attention)

Architectural innovation that drastically reduces KV cache memory usage (allowing huge context).

DeepSeek-R1

A reasoning model that outputs its "thought process" before the final answer.

Best Practices (2025)

Do:

• Use R1 for Math/Logic: It rivals o1-preview in math benchmarks.
• Local Distillations: Run DeepSeek-R1-Distill-Llama-70B locally for private reasoning.

Don't:

• Don't suppress thoughts: When using R1, the "thought" trace is valuable for debugging the model's logic.

References

• DeepSeek GitHub

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: G1Joshi
• Source: G1Joshi/Agent-Skills
• License: MIT
• Homepage: https://skills.sh/g1joshi/agent-skills

Install and usage instructions live in the source repository linked above.