After seeing DeerFlow hit #1 on GitHub Trending in late February 2026, I decided to install it locally on my Debian machine. I wanted a fully offline multi-agent system that can research, code, analyze data, and even generate podcasts — all powered by my own Ollama models. No API keys, no monthly bills, full privacy. Here’s exactly how I did it, including the pitfalls and fixes.
Hardware / Software context
- OS: Debian Trixie (testing)
- RAM: [your total RAM here, e.g. 16 GB — mention if you hit freezes with 14B model]
- CPU: [your CPU if relevant]
- GPU: none (CPU-only inference)
- Docker: official Docker Engine CE 29.3.0
- Ollama: running with several models pulled
Step-by-step installation log
- Fresh clone & config generationBash
cd ~ rm -rf deer-flow git clone https://github.com/bytedance/deer-flow.git cd deer-flow make config - Edited config.yaml to use local Ollama I replaced the example models section with my Qwen2.5 14B setup:YAML
models: - name: local-qwen display_name: Qwen2.5 14B (Ollama Local) use: langchain_openai:ChatOpenAI model: qwen2.5:14b-instruct api_key: dummy base_url: http://host.docker.internal:11434/v1 # key fix on Linux max_tokens: 8192 temperature: 0.3 supports_thinking: true(Tried 172.17.0.1 first, but host.docker.internal was the winner for container-to-host networking.) - Pulled the model in OllamaBash
ollama pull qwen2.5:14b-instruct(9 GB download — took ~20 minutes on my connection) - Launched DeerFlowBash
make docker-init # pulls sandbox image (once) make docker-start # starts everythingWaited 1–2 minutes → UI appeared at http://localhost:2026. - First successful test Created new chat → selected “Qwen2.5 14B (Ollama Local)” → sent: “Hello! Confirm you are Qwen2.5 14B Instruct running via Ollama in DeerFlow.”Response (paraphrased): “I confirm that I am Qwen-2.5-14B-Instruct […] I am currently running via Ollama in the DeerFlow environment. How can I assist you today?”Success! Full local LLM + agent harness.
Lessons learned / gotchas
- Memory matters — Qwen 14B + DeerFlow stack pushed my machine to swap hell and froze it once. Switched to smaller models (Llama 3.2, Gemma 2B) for stability.
- base_url networking — http://172.17.0.1:11434/v1 sometimes fails from containers. host.docker.internal worked reliably.
- 502 Bad Gateway — nginx couldn’t reach backend/gateway during startup. Waiting longer or restarting fixed it.
- Hydration mismatch warnings in console — harmless Next.js dev-mode noise. Ignore or test in incognito.
- Old threads 404 — after restart, old chat IDs disappear. Just start new chats.
Current verdict DeerFlow 2.0 is impressive: clean UI, powerful sub-agents, sandbox support, and it runs completely locally. Once I switched to a smaller model, it’s stable and fast. Already used it for quick Bayesian coin-flip analysis — worked flawlessly.
Repo: https://github.com/bytedance/deer-flow Official site: https://deerflow.tech
Have you tried it yet? Let me know in the comments what model you’re using and what tasks you’re throwing at it. https://deerflow.tech/