Ollama Model Library, By the Numbers

Originally published on DevToolHub.

Someone asks "what's the best Ollama model" in a Discord or a Reddit thread almost every day. The answer usually comes from memory, or a single favorite. So we pulled the actual data instead. Every model in the Ollama model library, its pull count, its capability tags, and when it was last updated. Current as of September 5, 2026 — the library changes weekly, so treat these counts as a snapshot, not a permanent ranking.

What's in the Ollama model library right now

The Ollama model library holds 239 models as of this pull. That spans general chat models, embeddings, vision models, and a growing "cloud" category that runs on Ollama's hosted infrastructure instead of your GPU. Combined, every model in the library has been pulled just over 1.04 billion times.

That total tells you less than it looks like it should. Pulls aren't spread evenly. A handful of general-purpose chat models account for most of it. The long tail of specialized or older models barely registers.

Where Ollama model library pulls actually concentrate

The top 10 models by pull count account for 56.4% of all pulls in the entire library. Meta's Llama family and DeepSeek's reasoning models lead by a wide margin:

Rank	Model	Pulls	Capabilities
1	llama3.1	119.2M	tools
2	deepseek-r1	92.4M	tools, thinking
3	nomic-embed-text	84.6M	embedding
4	llama3.2	82.4M	tools
5	gemma3	40.1M	vision
6	qwen2.5	39.3M	tools
7	qwen3	36.3M	tools, thinking
8	mistral	33.3M	tools
9	gemma2	31.8M	text only
10	llama3	25.2M	text only

Here's the thing about that list: it's mostly older releases. llama3.1 and llama3.2 alone outpull every newer Llama version combined. llama3 — several generations behind — still sits at 25.2M pulls. That's because pull counts accumulate over a model's lifetime. An older model with a long track record will always look bigger than a newer one that's objectively better. Don't read this table as "what to use today." Read it as "what people have been defaulting to."

nomic-embed-text at #3 breaks the chat-model pattern. It's a pure embedding model. Its position this high says a lot about how much RAG and semantic-search tooling runs on Ollama underneath the chat interfaces people actually see.

How many Ollama models actually support tool calling

If you're building an agent, this number matters more than pull count. Only 93 of 239 models (38.9%) carry the tools capability tag. The rest either don't support structured tool calling, or Ollama hasn't tagged them for it.

The full capability breakdown across the library:

• Tools (function calling): 93 models — 38.9%
• Thinking (extended reasoning): 42 models — 17.6%
• Vision: 38 models — 15.9%
• Cloud (hosted, not local): 18 models — 7.5%
• Embedding: 12 models — 5.0%
• No special tag (plain text completion): 118 models — 49.4%

These overlap — a model like gemma4 carries vision, tools, thinking, and cloud tags at once. But the practical read is simple. Half the library is plain text completion. If your use case needs tool calling specifically, you're choosing from under 40% of what's listed. Check a model's tags on its library page before you build around it, not after.

Why the Ollama model library carries 7,359 tag variants

Across all 239 models, there are 7,359 individual tags. Each one is a different quantization, parameter size, or context configuration of the same base model. That averages out to roughly 30.8 variants per model, though it's lopsided — llama3.1 alone carries 93 tags, while a narrow embedding model might have 3.

This ties into the same issue covered in the Ollama vs LM Studio comparison. Ollama's model names default to a specific quantization, usually Q4_K_M, a 4-bit quant, without making that choice obvious on the model card. With 30+ tagged variants per model on average, "the model" isn't one thing — it's a family. The tag you don't specify is a decision Ollama makes for you. If you're benchmarking against another tool, pin the exact tag (ollama pull llama3.1:8b-instruct-q8_0, for example) instead of the bare model name.

How much of the Ollama model library is still maintained

Freshness in the library skews old. Of the 239 models:

• 174 (72.8%) were last updated over a year ago
• 51 (21.3%) were updated sometime this year
• 9 (3.8%) were updated this month
• 5 (2.1%) were updated this week

Nearly three-quarters of the library hasn't been touched in over a year. That's not automatically a red flag. A model's weights don't need updates the way a CLI tool does, and a good 2024 model is often still a good model. But it does mean the library is mostly an archive with a small, actively-tended front section. If a model card shows no update in over a year, and a same-family successor has recent activity, the successor is usually the safer default.

What this means if you're picking a model to run

Two practical takeaways come out of this data. First, pull count is a popularity signal, not a quality signal. It rewards models that have been available the longest. That's why year-old Llama releases still outrank newer, often better models. Cross-check pull count against a model's actual release date before treating it as a recommendation.

Second, the capability tags are the fastest filter for narrowing 239 models down to the ones that fit your use case. Building an agent? Filter to the 93 tools-tagged models first. Need a local embedding model for RAG? nomic-embed-text and mxbai-embed-large are the two with real usage behind them. For everything else — running the model once you've picked it — the hardware requirements guide and the Ollama API guide cover the setup.

Frequently Asked Questions

Q: How many models are in the Ollama library? A: 239 models as of September 5, 2026. The Ollama model library is updated regularly, so this number changes — check ollama.com/library directly for the current count.

Q: What is the most popular model in the Ollama library? A: llama3.1, with 119.2 million pulls, ahead of deepseek-r1 at 92.4 million and nomic-embed-text at 84.6 million. All three are well over a year old, which is typical — pull counts accumulate over time and favor established models.

Q: How many Ollama models support function calling? A: 93 of 239 models (38.9%) carry the tools capability tag, which indicates support for structured function calling. Check the individual model's page on ollama.com to confirm before building an agent around it.

Q: Why does Ollama have so many tags for one model? A: Each tag is a different quantization, parameter size, or configuration of the same base model. The library averages about 30.8 tags per model, and the untagged default usually points to a 4-bit quantization rather than the highest-quality version available.

Quick Summary:

• 239 models in the Ollama model library as of September 5, 2026, with 1.04 billion cumulative pulls.
• The top 10 models account for 56.4% of all pulls — usage concentrates hard in a handful of established releases.
• Only 38.9% of models carry the tools (function-calling) tag; 49.4% have no special capability tag at all.
• The library averages 30.8 tag variants per model, so the untagged default is a quantization choice you're making without realizing it.
• 72.8% of models haven't been updated in over a year — the library is mostly an archive with a small active front section.

If you've picked a model from this list and need to know whether your hardware can run it, the Ollama hardware requirements guide breaks down RAM and VRAM by model size. And if local hardware runs out before your model does, the Ollama Cloud pricing and limits guide covers the hosted tier.