How to increase Ollama's context length (num_ctx)

Ollama starts models with a short context window, and when a prompt is longer it quietly drops the start of it. If a model "forgets" the beginning of a document you pasted, or ignores your instructions halfway through a long chat, this is usually why. You can raise it three ways: for everything with an environment variable, for one session in the chat, or permanently for one model.

The setting is called num_ctx, measured in tokens. A token is roughly three quarters of an English word.

Option 1: For every model (environment variable)

Set OLLAMA_CONTEXT_LENGTH where the Ollama server runs, then restart Ollama:

That's the quick test. To make it stick, set it the same way you'd set OLLAMA_HOST: in the systemd service on Linux, with launchctl setenv on a Mac, or in the system environment variables on Windows. Ollama on your network has the steps for each. Recent versions of the desktop app also have a context length setting in Settings.

This is the best option if you use Ollama through other apps, including ones that use the OpenAI-compatible /v1 endpoint, which can't set context per request.

Option 2: For one chat (in the REPL)

Inside ollama run, type:

/set parameter num_ctx 16384

It applies until you exit. To keep it, save a copy of the model under a new name with /save qwen3-16k, then ollama run qwen3-16k next time.

Option 3: For one model (Modelfile or API)

Make a variant with a longer context baked in. Create a file named Modelfile:

FROM qwen3:14b
PARAMETER num_ctx 32768

Then build it:

This doesn't download or copy the model again; it just stores the new settings. More on Modelfiles in importing a GGUF into Ollama.

From the API, pass it per request on the native endpoints:

See the Ollama API guide for the rest.

Check it worked

Load the model, then run:

Recent versions show the context each loaded model is using. Also watch the processor column. If it changed from "100% GPU" to a CPU/GPU split after you raised the context, the extra memory pushed part of the model off your card, and it will be much slower.

What it costs

Context lives in the KV cache, which sits in graphics memory next to the model. For a model like Llama 3 8B it's about 128 KB per token: 8K tokens adds about 1 GB, and 32K adds about 4 GB. Bigger models need more per token.

Don't go past what the model was trained for. From our model pages: Qwen3 8B and Qwen3 14B handle 32K natively, Gemma 3 12B and gpt-oss 20B handle 128K.

Leave room on your card

Our calculator's fit check assumes about 1.2 GB for context on top of the model file. Here's a 16 GB RTX 4060 Ti:

On that card, Qwen3 14B leaves about 5 GB spare, so 16K or even 32K of context fits. gpt-oss 20B leaves very little. If you need lots of context, pick a model a size down rather than spilling onto the CPU.

Tips

• Raise it only as far as you need. A bigger window costs memory even when it's empty, and prompt processing slows as it fills.
• Running several requests at once multiplies it. With OLLAMA_NUM_PARALLEL above 1, each slot gets its own context.
• Short on memory? Setting OLLAMA_FLASH_ATTENTION=1 and OLLAMA_KV_CACHE_TYPE=q8_0 roughly halves the cache's size with little quality loss. See fixing out-of-memory errors.

For the theory behind all this, read context length and the KV cache.