Announcing support for DeepSeek-R1 in our IDE plugin, self-hosted by Qodo

Since the release of OpenAI o1 (and the impending release of o3), it has been regarded as the leading LLM with the ability to reason and reflect, bringing AI closer to System 2 thinking. While models like Anthropic’s Sonnet 3.5 are very good at everyday coding tasks, for more complex coding challenges, the developers at Qodo tend to choose o1. (Read our opinionated comparison of leading LLMs) The differing capabilities of models is why we make all the top tier models available to use in Qodo Gen, Qodo’s AI-powered coding assistant. 

Recently announced, DeepSeek-R1, the newest frontier LLM from the minds at DeepSeek, is a reasoning model on par with o1, demonstrating that open source models can be competitive in this domain. Across social media and forums, DeepSeek-R1 is being hailed as a gamechanger:

We’ve built our platform to easily enable support for any new model, and we’re happy to announce that Qodo is self-hosting DeepSeek-R1-Distill-Qwen32B and have added it to Qodo Gen, giving our users safe and secure access to R1 directly within their IDEs and as part of their coding workflows. 

The rise of “thinking models”: what sets OpenAI o1 and DeepSeek-R1 apart

Inference models like Anthropic’s Claude Sonnet 3.5 and GPT-4o primarily interpret and respond to input. These models are reactive—giving answers based on patterns and probabilities learned during training. By leveraging technologies like Retrieval-Augmented Generation, we can make these models context-aware, but they have less depth in reasoning and may not always handle multi-step problem solving effectively. 

OpenAI o1 and DeepSeek-R1 utilize reinforcement learning to approach tasks with reasoning rather than simply predicting the most likely outcome. By leveraging test-time computation, these models perform iterative processing, enabling them to tackle more complex tasks. These models handle problems that require deduction, logic, reflection, or extrapolation beyond observed patterns. For example, they might analyze a large codebase, suggest architectural improvements, or debug intricate errors by reasoning about cause and effect.

The flip side of high reasoning: cluttered chain of thought 

In our initial trials, we found R1 to be adept at identifying when a prompt is ambiguous or unclear, reflecting on the gaps in the prompt and asking the user for more information: 

While this reflective behavior is a hallmark of advanced reasoning models, its chain of thought can be noisy—a characteristic that can be both helpful and overwhelming. Its approach mirrors humanlike brainstorming, where almost every detail or path is examined thoroughly. While this allows the model to cover a wide range of considerations, it can lead to verbose explanations. 

How does R1 compare to o1? 

Democratizing access to AI 

The most important distinction between OpenAI o1 and DeepSeek-R1 is that R1 is open-source. Traditionally, high-performing models have been locked behind proprietary APIs and commercial terms, limiting broader community involvement. By contrast, R1 enables users to explore, modify, and deploy the model without constraints.

Cost efficiency

One of the most notable advantages of DeepSeek-R1 is its efficiency and speed. DeepSeek reports that R1 runs at a fraction of the cost compared to leading proprietary models—up to 30x cheaper—and can generate responses 5x faster.

Performance 

DeepSeek-R1 is standing toe-to-toe with o1 across several benchmarks, edging it out on AIME 2024 (79.8% Pass@1) and matching o1’s performance on MATH-500 with 97.3% score. While it scores slightly below o1 on certain knowledge benchmarks such as MMLU, DeepSeek-R1 still surpasses most other closed-source models. It has expert coding abilities—evident in a 2,029 Elo rating on Codeforces—further highlighting that R1 is a serious contender to o1 for both complex reasoning and real-world development tasks.

Getting started with DeepSeek-R1 in Qodo Gen

It’s easy to start using DeepSeek-R1 with Qodo Gen using our IDE plugins for VSCode and Jetbrains. Once installed, you can select DeepSeek-R1 from our library of supported models and use it to generate code and tests.