Ollama Shipped More Than 20 Models in Two Weeks; Picking One Is Harder Than Running One

Ollama posted an updated list on X of the models from the past two weeks. Six on cloud, six local, plus a string of decision models—11 by count. The official account later added in a reply that they forgot to include LFM3.

Let's start with the list.

Cloud models include:

Local models include:

Decision models are listed separately as a set: Clef-flash, D1, D1-omni, D3, Kev, Laya, Nimble, Nimble-v3, Pplx-decider, Strands-decider, Tev1.

This part is more worth looking at than the model names. According to the descriptions on Ollama's page, Clef turns state and a set of typed question schemas into decisions; D3 turns text, structured data, and images into scored choices, yes/no answers, and ratings; Nimble-v3 is used for fast classification, routing, and scoring. They are not for chatting; they are for making choices, routing, and judgments inside a call chain.

Eleven decision models appearing in two weeks is a direct signal: now that there are more models, how to choose a model has itself become a layer of work.

One user wrote in the replies that with 11 decision models, they still do not know which one to download. That comment is quite representative. Ollama's search page now has filters such as Local, Cloud, Tools, Thinking, Vision, Embedding, and Decision. The model list keeps getting longer, and what ordinary users see first is not capability but the cost of choosing.

The LFM3 that the official account added later also deserves separate mention. LFM3-1.7B is a text model based on a tri-hybrid architecture, aimed at local agents and specialized applications. It has a 256K context and takes about 1GB of memory at 4-bit precision. The page states clearly that it is suitable for local agents, tool calling, structured output, and single-GPU fine-tuning, and is not recommended for agentic coding or complex multi-step reasoning.

The LFM3 page also lists launch methods for agent harnesses such as Claude Code, OpenCode, Hermes Agent, and OpenClaw. Small models paired with local agents—this direction is more practical than parameter races.

The focus in the comments is also very practical. Someone asked which small model on a Mac mini is best for tool calling, concerned about whether it can still hold the JSON schema by step 6. Someone asked how much RAM each local model needs. Someone said Ollama caps Qwen's default context at 32k even though training used 262k, and you only find out by digging through logs. Someone else asked when Phonon will support Linux and how it performs with low VRAM.

These questions affect daily use more than leaderboard rankings. Whether a model can run is one thing; default context, VRAM usage, and whether structured output is stable are another.

With this update, Ollama put cloud and local models on the same list and separately pulled out decision models. The direction is already clear: it is not just a tool for running models locally, but is also becoming a model gateway. As the gateway gets longer, choosing becomes the new problem.

Model list here: https://ollama.com/search

LFM3 page here: https://ollama.com/library/lfm3

If you want to try it, install Ollama and run:

ollama run lfm3