Skip to content
localmodel.run

Guide · Local coding

Can you run Claude locally?

Anthropic does not offer Claude weights for local download in its published model catalog. You can use the Claude Code CLI with a separately hosted local model through Ollama. Claude Code is the coding tool; the selected model supplies the responses.

The short version

Claude is available through Anthropic's cloud/API catalog, not as published local weights. Ollama's Claude Code integration lets the Claude Code harness select a local model. That is useful for a local workflow, but it is not a local Claude release and does not promise Claude-equivalent quality, behavior, or tools.

What is and is not local

Anthropic's model overview lists Claude as a cloud/API offering. It does not publish the model weights or a GGUF download, so there is no Claude file to put on your Mac, PC, or phone for offline inference. This answer was checked on 12 September 2026.

Claude Code is a coding environment that works in the terminal, IDEs, desktop app, and web. Its official overview also documents third-party providers for the CLI. A local provider changes where the selected model runs. It does not turn that model into Claude.

Use Claude Code with Ollama

Install Ollama and Claude Code, then follow Ollama's integration guide. Launch with: 

ollama launch claude

The guide lets you choose local or cloud models. Ollama uses its local endpoint on port 11434 by default and recommends 64K or more context for large repositories. Choose a local model that fits your hardware, then check its own documentation and licenses. Some external tools or services used by a coding workflow may still need network access.

Memory for local coding models

These examples show memory requirements; we have not tested these models with Claude Code. The totals use this site's Q4_K_M memory estimate: measured weights where available, plus an estimated KV cache and 0.8 GB runtime overhead. A 64K context can be valuable for a repository, but it costs much more memory. Use the calculator for your device and the coding leaderboard for sourced benchmark coverage.

Estimated Q4_K_M memory at 4K and 64K context for three coding models.
Local coding model Q4_K_M weights Est. total, 4K Est. total, 64K
Qwen2.5 Coder 7B 4.36 GB ~5.8 GB ~15.4 GB
Qwen2.5 Coder 14B 8.37 GB ~10.1 GB ~23.6 GB
Qwen3-Coder 30B-A3B 17.28 GB ~19.4 GB ~39.3 GB

Memory-data refresh: 2026-09-14. These are model-memory estimates, not total system RAM requirements; leave room for the operating system, coding tools and other apps. See the memory methodology for assumptions. Larger context also needs support and configuration in the model and runtime.

FAQ

Can you run Anthropic Claude locally?

No. Anthropic's published Claude model catalog describes cloud/API models, and Anthropic has not published Claude weights or GGUF files for local inference. A local model used through Claude Code is a different model.

Can Claude Code use a local model?

Yes. Claude Code is the app and coding harness, not the model weights. Ollama documents its Claude Code integration, where you can choose a locally hosted model. That does not make the local model Claude or establish matching quality or tool behavior.

How much memory does a local coding model need?

It depends on the model and context window. In the estimates below, Qwen2.5 Coder 7B needs about 5.8 GB at 4K context and 15.4 GB at 64K. The weights are only part of the total; the KV cache grows with context.

Weighing a local model against a hosted Claude API? Compare local models, Claude, Astra and Gemini using preference ratings, your hardware's memory and monthly API token costs.

Sources