MLNLP community is a well-known machine learning and natural language processing community both domestically and internationally, covering audiences including NLP graduate students, university professors, and researchers from enterprises.

Community Vision is to promote communication and progress between the academic and industrial sectors of natural language processing and machine learning, especially for beginners.

Source | Quantum Bit

Author｜Mengchen Xifeng

Let DeepSeek replace Claude’s thinking; the integration approach has become popular.

The reason is simple: it performs better than using DeepSeek R1, Claude Sonnet 3.5, or OpenAI o1 models individually.

First, let’s look at a VCR:

Next, let’s examine a review result:

In the code editing benchmark Polyglot Benchmark, the integrated model slightly outperformed o1-high and R1.

In this test, R1 acted as the architect, describing how to solve code issues.

While Claude acted as the programmer, generating specific code editing instructions as required to apply changes to the source files.

Additionally, several interesting conclusions were drawn during the experiment:

• The combination of o1 and Claude Sonnet does not perform as well as using o1 alone.

• Using R1 or o1 as the architect, with other models besides Claude as the programmer, does not perform as well as using R1 or o1 alone.

• However, using o1-preview and o1-mini as architects, with many different models as programmers, can improve the combined performance.

• The reasoning process token of R1 is less effective than the final output token of R1.

It seems that R1 and Claude Sonnet really are a perfect match!

The DeepClaude application itself is 100% free and open-source, having gained 3k stars on GitHub (of course, you need to use your own API).

After testing, netizens summarized that Claude excels at writing clear and well-structured text and code, allowing it to transform DeepSeek-R1’s ideas into concise replies.

One of the authors of DeepClaude expressed:

AI agents and agent applications are showcasing a paradigm shift towards a “digital world first,” where intelligent systems are becoming proactive collaborators rather than just passive tools.

Hybrid of DeepSeek and Claude

Specifically, DeepClaude is an LLM inference API written in Rust.

It provides a unified interface that seamlessly connects the CoT logical reasoning capabilities of DeepSeek R1 with the replies of Claude in a single flow.

Developers can use this API to simultaneously call the functionalities of both models while maintaining full control over their API keys and data.

The team that created it is called Asterisk, consisting of members with backgrounds in security research & CTF (Capture The Flag), dedicated to using AI to make code security checks more efficient.

The team believes that the CoT deep reasoning of DeepSeek R1 has even reached a level of metacognition, enabling it to self-correct, think through uncommon/extreme/special situations, and perform reasoning similar to Monte Carlo Tree Search (MCTS) in natural language.

However, R1 lacks in code generation, creativity, and conversational skills, while Claude 3.5 Sonnet excels in these areas, making it a perfect complement.

Why not combine the two? Take the strengths of both to create DeepClaude!

In conversations, before Claude responds, the system displays pre-filled text like ““.

DeepClaude combines these two models, featuring the following characteristics:

The hosted API is completely free, allowing users to use their own keys and integrate the streaming APIs of DeepSeek and Claude, providing conveniences such as calculating combined usage and pricing.

The code is open-source, allowing users to host, modify, and redistribute freely. The team states it has been used on a large scale in Asterisk’s production environment, handling millions of tokens daily without any failures, as long as there is no abuse.

One More Thing

Did you think that combining two models is the limit?

No no no

Netizens have also developed a three-model integration approach, combining the thinking results of DeepSeek-R1 and Gemini 2.0 Flash, allowing Claude Sonnet to answer questions.

They also achieved good results in the GPQA test (PhD-level multiple-choice questions in physics and chemistry that Google cannot find).

GitHub link: https://github.com/getasterisk/deepclaude

References: [1]https://aider.chat/2025/01/24/r1-sonnet.html[2]https://x.com/deepclaude_/status/1886911416478642279[3]https://x.com/omercelik/status/1883510797193937278[4]https://x.com/mufeedvh/status/1883620781583901011

Technical Community Invitation

△Long press to add assistant

Scan the QR code to add the assistant on WeChat

Please note: Name-School/Company-Research Direction

(e.g., Xiao Zhang-Harbin Institute of Technology-Dialogue System)

to apply to join the Natural Language Processing/Pytorch technical community

About Us

MLNLP Community is a grassroots academic community jointly built by machine learning and natural language processing scholars from home and abroad, which has developed into a well-known community in machine learning and natural language processing, aiming to promote progress among the academic and industrial sectors as well as enthusiasts.

The community can provide an open communication platform for related practitioners’ further education, employment, and research. Everyone is welcome to follow and join us.