Ollama default 262144 context and num_ctx memory cost on Mac
Parent: Mac local LLMs: Ollama internals · Published reference · snapshot 2026-10-05
↓ Facts as markdownall context files
Ollama docs: "Tasks which require large context like web search, agents, and coding tools should be set to at least 64000 tokens."
These notes link each claim to its source. Some citations identify archived research reports that are no longer hosted on this site. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.
Facts
- Ollama docs: "Tasks which require large context like web search, agents, and coding tools should be set to at least 64000 tokens." [source]
- Ollama docs: cloud models are set to their maximum context length by default. [source]
- Ollama docs example: `ollama ps` columns are NAME, ID, SIZE, PROCESSOR, CONTEXT, UNTIL, e.g. gemma4:latest 9.6 GB, 100% GPU, context 131072. [source]
- Ollama docs advise using the model's maximum context for best performance while keeping PROCESSOR at 100% GPU (no CPU offload). [source]
- Per-token KV cost differs by model: a maintainer measured 32k context at 6.2 GB for phi4:14b and 12 GB for gemma3:12b (gemma3 also needs memory for its image projector). [source]
- A maintainer showed OLLAMA_NUM_PARALLEL=3 turned a 128000 request into --ctx-size 384000, needing 181.6 GiB KV and leaving only 12 of 63 layers on GPU. [source]
- Changing num_ctx causes a model reload, which is why the first request after a change is slow. [source]
- If a model is sharded across several devices the VRAM needed rises because each device runs a copy of the compute graph. [source]
Children
- No children recorded.