Point the ollama create command at a Modelfile and it builds you a custom local model — your own base model, system prompt and settings bundled into something you can run with a single ollama run call. And as of Ollama v0.34.1, released 14 September 2026 per the official GitHub release notes, create is no longer experimental for MLX models, which means Apple Silicon users finally get first-class custom model building instead of a feature hidden behind a warning label.
📺 Watch: Here's How to Run AI Models Locally!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →
What the ollama create command actually does
Ollama create takes a plain-text recipe called a Modelfile and turns it into a named model in your local library. The official Ollama CLI documentation shows the flow in two steps: first you write a Modelfile, then you run ollama create with the -f flag pointing at it. The Modelfile itself can be as simple as two lines — a FROM line naming a base model such as gemma4, and a SYSTEM line with the instructions you want baked in permanently.
That sounds small, but it changes how you work with local models. Instead of pasting the same system prompt into every session, you bake your instructions, persona and defaults into the model itself. Your customised model then shows up in ollama ls alongside everything you have pulled from the registry, and any tool that talks to Ollama — a chat interface, an agent framework, an editor plugin — can call it by name. If you have been comparing local chat setups, the Atomic Chat vs Ollama breakdown covers where Ollama's model management approach wins.
What changed in Ollama v0.34.1 on 14 September 2026
The headline change in v0.34.1, quoting the release notes directly, is that "the ollama create command is no longer experimental for MLX models". MLX is the machine-learning framework built for Apple Silicon, so this is aimed squarely at Mac users: creating custom models from MLX safetensors is now a supported, stable path rather than an experiment that might break between releases.
The same release note adds an important counterweight: GGUF model creation now mandates llama.cpp tooling for safetensor conversion and quantisation. In plain terms, Ollama has split the two model formats onto two roads. MLX safetensors go through ollama create natively. GGUF files — the format most quantised community models ship in — now lean on llama.cpp's own conversion and quantisation tools before Ollama picks them up.
Beyond create itself, v0.34.1 shipped performance work that matters for anyone running a large local library. The release notes report the /api/tags endpoint dropping from 3.1 seconds to 294 milliseconds on large libraries, which is the difference between a model picker that stutters and one that feels instant. Memory handling for MLX on Apple Silicon was improved, the repeat-token detection threshold was raised to 100 tokens to cut false positives in OCR-style workloads, and the typical_p sampling parameter was deprecated for new models while existing GGUF models keep support. A v0.34.2 pre-release followed on 15 September 2026 with further llama.cpp updates.
If you want a local AI stack that actually earns its keep — custom models feeding real automations instead of demos — the AI Profit Boardroom is where the full systems live: check out the AI Profit Boardroom. Prefer to talk it through 1-on-1 first? Book a free SEO strategy session and map your setup with Julian.
How to use the ollama create command step by step
The workflow the official docs describe is short enough to memorise:
- Write a Modelfile. A plain text file, no special extension needed. Start with FROM plus the base model you want to build on — the docs use gemma4 as the example — then add a SYSTEM line containing the instructions your model should always follow.
- Run create. Call ollama create with the -f flag pointing at your Modelfile. Ollama assembles the new model and registers it in your local library.
- Run it. Start the model with ollama run just like any pulled model, check it appears with ollama ls, and monitor it with ollama ps while it is loaded.
Because the created model is a normal Ollama model, everything downstream just works: you can point an agent at it, expose it through Ollama's API, or set it as the default brain for a local assistant. The practical question is which base model deserves your Modelfile, and that depends on your hardware and the job — the best Ollama model for Hermes Agent guide ranks the realistic options rather than the leaderboard bait.
📺 Watch: How to Use Local AI Models With Agent OS
The GGUF caveat you should not skip
The v0.34.1 change is genuinely two changes, and the second one will catch people. If your custom model workflow starts from safetensors that you previously converted and quantised inside Ollama's tooling for GGUF output, that path now runs through llama.cpp's tools by mandate. The release notes are explicit that GGUF creation "now mandates using llama.cpp tooling for safetensor conversion and quantization".
What does that mean for you in practice? If you are on Apple Silicon, the clean answer is to stay on the MLX path — it is the one Ollama just stabilised, and it is the one getting memory-handling improvements. If you are on other hardware where GGUF remains the default format, budget a little extra setup time for llama.cpp's conversion utilities. Neither road is closed; they are just clearly signposted now, and pretending the split does not exist is how you lose an evening to a failed conversion.
📺 Watch: Hermes Desktop Can Now Set Up Local AI in ONE Click
Why this matters for Hermes Agent and local automation
A custom model is only as valuable as the work you route through it. The pattern that pays is building a model with your operating instructions baked in, then letting an agent drive it around the clock. That is exactly the setup covered in the how to setup Hermes with Ollama walkthrough — Hermes as the agent layer, Ollama serving the model underneath — and the reasons given there for going local in the first place (cost, privacy, reliability) all get stronger when the model is one you shaped yourself with ollama create.
The same logic extends to a full local stack. The Ollama Hermes guide covers wiring the two together in detail, and the DeepSeek V4 Ollama write-up shows the same pattern with DeepSeek's open weights as the base — a natural candidate for a FROM line in your next Modelfile. If you want the broader system this plugs into, the Agent OS guide explains the operating layer that turns a single custom model into a business workflow, and the Goldie Bench write-up covers how the different model brains compare in hands-on tests, so you can pick your base model on evidence rather than vibes.
Is the ollama create command free, and should you use it?
Ollama is free and open source, and the create command ships with it — there is no paid tier gating custom model creation. Your only costs are hardware and time. On a recent Mac, the MLX path stabilised in v0.34.1 makes Apple Silicon arguably the most friction-free place to build custom local models right now: supported creation from safetensors, improved memory handling, and a fast library API for whatever front-end you put on top.
Should you use it? If you run the same system prompt into a local model more than twice a week, yes — baking it in with the ollama create command costs five minutes and removes a whole category of copy-paste errors. If you are still choosing your stack, start with the setup guides linked above, get one base model running well, and only then create your customised version. Custom models compound the value of a working setup; they do not rescue a broken one.
Quick answers on ollama create
Does ollama create cost anything? No — it ships with the free, open-source Ollama install, on every platform Ollama supports.
Do you need it just to run models? No. Pulling a ready-made model with ollama pull and starting it with ollama run needs no Modelfile at all. Create is for when you want your own instructions, persona or defaults baked into a named model.
What actually changed on 14 September 2026? Per the v0.34.1 release notes: create left experimental status for MLX models, GGUF creation now requires llama.cpp tooling for conversion and quantisation, the /api/tags endpoint got roughly ten times faster on large libraries, and the typical_p parameter was deprecated for new models while existing GGUF models keep support.
Where does a created model live? In your local Ollama library, next to your pulled models — list it with ollama ls, watch it while loaded with ollama ps, stop it with ollama stop, and delete it with ollama rm when you are done with an experiment.
If you want the shortcut — working Modelfiles, local agent configs and the exact stack behind these builds — it is all inside the AI Profit Boardroom: get the Agent OS and the full local AI system here. And if you would rather have a human walk your plan first, book a free SEO strategy session.
