Best Ollama models: pick one by task and memory, not by rank

Start with the capability tags under the model name. Tools means the model is tagged for tool calling, vision means it accepts images, thinking means it has a reasoning mode, embedding means it returns vectors rather than text, audio means audio features, and cloud means a tag runs on Ollama's servers instead of your machine. A tag describes a family, not every size in it, so check the tag table for the size you will pull.

The tag table gives a name such as gemma4:e4b, a download size, a context window and the input types. The name after the colon picks the size, and the default tag, latest, points at one of them. Pull counts and update ages on the library page measure popularity and recency, not quality.

A tag ending in cloud, such as gemma4:31b-cloud, is not local. Your prompts go to Ollama's servers. See Ollama Cloud pricing for what that means.

Ollama's own documentation gives no formula for weights and context together, so this page states none of its own. What the sources do say is this. The tag size is the download. Google's Gemma 4 table gives approximate memory to load the weights, including 20 per cent overhead, and says it excludes the context cache: at 4-bit, 17.5 GB for the 31B model and 2.9 GB for E2B. A larger context window needs more on top.

Ollama's context length page says the default context depends on video memory: 4k below 24 GiB, 32k from 24 to 48 GiB, and 256k at 48 GiB or more. It says tasks such as web search, agents and coding tools should use at least 64000 tokens, and that more context needs more memory. The FAQ still says the default is 4096 tokens, so set the value you need explicitly. After loading, run ollama ps and read the PROCESSOR column to see whether the model sits fully on the GPU.

Ollama's FAQ also describes a quantised key and value cache. OLLAMA_KV_CACHE_TYPE can be q8_0, which it says uses about half the memory of f16, or q4_0, about a quarter. It needs Flash Attention, and the effect on quality depends on the model.

The Ollama software is MIT-licensed, but every model carries its own licence, and the licences in the task table are not the same. Most of the ten show Apache License 2.0. Llama 3.2 shows a community licence agreement with an acceptable use policy. DeepSeek R1 shows MIT but notes that distilled models derive from other base models. embeddinggemma shows the Gemma Terms of Use, while Gemma 4 shows Apache 2.0, so two models from one family can carry different terms.

Read the licence file on the tag page you pull, rather than the family page alone. Licences can change between versions, and fine-tunes and community uploads carry their own. This page summarises what the pages displayed on 2026-10-07. It is not legal advice, so have counsel read the text before you ship a model inside a product. A cloud tag is also a different service with its own terms.

You do not need Swfte to pull and run an Ollama model, and Swfte does not choose a model for you or rank them. Where several people or services share a runtime, Swfte Connect gives you one OpenAI-compatible API with bring-your-own-key, per-workspace routing rules, budgets and usage caps, content-policy detectors for secrets and personal data, and an audit event stream. These are Built. It lets you change which model a workload uses as a routing decision instead of a setting inside each client. A hosted gateway needs a network path to your Ollama server.

Swfte Cortex is a desktop app that uses local models through Ollama and LM Studio. That is Built. See Connect for the gateway.