DeepReasoning — Chain-of-Thought Inference API Merging DeepSeek R1 and Claude

Introduction

DeepReasoning is a Rust-based inference proxy that combines DeepSeek R1 chain-of-thought reasoning with Anthropic Claude model outputs. It exposes an OpenAI-compatible API and a chat UI, enabling applications to benefit from explicit reasoning traces while leveraging Claude for final responses.

What DeepReasoning Does

• Proxies inference requests through DeepSeek R1 for chain-of-thought reasoning and Claude for final output
• Exposes an OpenAI-compatible REST API for drop-in integration
• Provides a web-based chat UI that displays reasoning traces alongside responses
• Streams both reasoning steps and final answers in real-time
• Supports configurable model selection for each stage of the pipeline

Architecture Overview

DeepReasoning runs a Rust HTTP server built on Axum. Incoming requests are first sent to DeepSeek R1 to generate a chain-of-thought reasoning trace. The trace is then included as context in a follow-up request to Claude, which produces the final response. Both stages stream tokens to the client as they arrive. The chat UI is a bundled frontend that renders reasoning traces in collapsible sections.

Self-Hosting & Configuration

• Clone and build with Cargo (requires Rust 1.70+)
• Set DEEPSEEK_API_KEY and ANTHROPIC_API_KEY as environment variables
• Configure model versions, temperature, and max tokens via config.toml
• Runs on localhost by default; configurable port and host binding
• Docker image available for containerized deployment

Key Features

• Combines chain-of-thought reasoning from DeepSeek R1 with Claude output quality
• OpenAI-compatible API allows integration with existing tools and libraries
• Rust implementation handles concurrent requests with minimal resource usage
• Web UI displays reasoning traces for transparency and debugging
• Supports streaming for both reasoning and response phases

Comparison with Similar Tools

• OpenRouter — multi-model API gateway; DeepReasoning specifically chains reasoning and response across two models
• LiteLLM — unified LLM proxy; DeepReasoning adds a two-stage reasoning pipeline
• Portkey — LLM gateway with caching; DeepReasoning focuses on CoT integration rather than routing
• Jan — offline AI desktop app; DeepReasoning is an API server for programmatic access
• LobeChat — multi-model chat UI; DeepReasoning provides a specialized reasoning-chain interface

FAQ

Q: Do I need both DeepSeek and Anthropic API keys? A: Yes. The two-stage pipeline requires access to both providers.

Q: Can I use other models instead of DeepSeek R1 or Claude? A: The architecture supports any OpenAI-compatible reasoning model for stage one and any Anthropic model for stage two. Custom providers can be configured.

Q: What is the latency overhead of the two-stage approach? A: Total latency is roughly the sum of both API calls, but streaming mitigates perceived delay since reasoning tokens appear immediately.

Q: Is this suitable for production use? A: The Rust server is production-grade in terms of performance and reliability. Cost and latency depend on the underlying API providers.

Sources

• https://github.com/winfunc/deepreasoning