The best open source LLM for practical work is the one whose license you can accept, whose size fits the hardware, and whose card says it can do the job. This page does not name a benchmark winner, and it does not invent scores for any Llama version or any other family. Model names change quickly. The model card wins over a roundup, including this one. When you are ready to run the file, use Best Local LLM Tools to Run on Your PC.
Key takeaway: Read the license, compare size with the machine, check whether tool use is documented, and test the model on your task. A leaderboard screenshot is not a selection.
Research reviewed: October 30, 2026. This is desk research by the AI Savr editorial team. We did not run a private lab test, and we did not invent scores, traffic numbers, or hands-on results. Confirm live prices on the vendor page before you buy.
How do you choose without a winner?
Treat "open" as a question, not a badge. A downloadable weight file is not the same as a license that lets you ship a product, fine-tune, or keep outputs. You will see family names such as Llama, Mistral, Qwen, and Gemma. Those are labels to look up, not a ranking. Whatever version number is current this month, open that card and read it. Ignore a roundup that crowns any of them with a score you cannot find on the card.
| Question | Where to look | Ignore this |
|---|---|---|
| License | The model card | A roundup ranking |
| Size versus hardware | The card and a local test | A benchmark crown |
| Tool use | The card's tool notes | A blog score |
| Does it do our job | A small test on your task | A leaderboard screenshot |
Four places the decision actually happens
Hugging Face for the model card
Hugging Face is the first stop because the card is the document that beats a roundup. Read the license, the size, the language or tool notes, and the exact file you would download. Family names change, and two files with similar names can have different terms. If the card is missing, you do not have enough to choose.
Choose it when: you are about to recommend a model to anyone else.
Leave it when: you want a chat window before you know the license. That order is backwards.
Price and access: Reading a card is free. Some weights ask you to accept terms before download. Confirm that gate on the card. This guide does not quote a paid hub plan.
A real use: Open two cards in the same family and write down license, size, and whether tool use is documented. If one card forbids your use, drop that file even if a blog ranked it first.
llama.cpp for size versus the machine
llama.cpp is how you find out whether the file fits the hardware you actually have. The card's size is the claim. The run is the check. A model that does not fit is the wrong model for that PC, no matter what a leaderboard said. This project does not declare a winner.
Choose it when: you need to know if a candidate will run on a given machine.
Leave it when: you have not picked a candidate from a card yet.
Price and access: Read the LICENSE file in the repository before you embed it in a product. Do not treat a popular GitHub project as permission for every commercial use.
A real use: Load the smallest acceptable file from the card and run a prompt from your real task. If the machine stalls, stop. Choose a smaller file or a different machine. Do not cite a public score as a reason to force it.
Ollama for a careful local try
Ollama is a practical way to try a model you already chose, on a machine you control. It is not a catalog of winners. Pull the file that matches the card, not whichever name is featured in the app this week. Remember that a local app is not private if it phones home or offers a cloud model.
Choose it when: the card is acceptable and you want a quick task test on a PC.
Leave it when: the license is unclear. Do not pull first and read later.
Price and access: Confirm the license and any cloud option on ollama.com. A convenient install is not a model evaluation.
A real use: After the card check, pull that specific model and give it ten examples from the real job. Score them yourself against a written rubric. Do not replace the rubric with a public benchmark.
Claude for the task check, not as an open model
Claude is not an open-source model, and it is not on this list as a winner. It is a useful comparison seat after a local model produces an answer. Run the same task on your notes and see which output a person can use. If the hosted model is better, say so, and also say what data you are willing to send it.
Choose it when: you need a practical comparison on your task, not a public scoreboard.
Leave it when: the data cannot be pasted into a hosted tool. Stay on the local file.
Price and access: Pro is $20 a month, or about $17 on the annual plan, on Anthropic's pricing page as of late October 2026. Max starts from $100. Confirm before you buy.
A real use: Give Claude and the local model the same five redacted tasks. A person marks which answers they would ship. Keep the winner of that sheet, not the winner of a blog chart.
Select a model in five checks
Selection is a short record you can show a colleague. It is not a screenshot of someone else's chart.
- Write the job. Summaries, a local coding helper, or extraction. A model can be fine at one and useless at another.
- Read the license on the card. Decide if you may use, modify, and ship it. A download button is not a license.
- Compare size with the hardware. If it does not fit, it is not a candidate for that machine.
- Read the tool-use notes. If the job is calling functions or editing a repo, a chatty model with no tool notes is a weak bet.
- Test on your examples. Ten real tasks, scored by a person. Then write down the file name, because the marketing name will change.
What a roundup cannot know for you
Model names change quickly, and a version that was reasonable last month may have a new license or a new size. This page will not be updated every time a lab renames a file. Open weight is not the same thing as open source, and neither phrase tells you the license. No score for Llama, Mistral, Qwen, Gemma, or any other family appears here, because a score you did not measure is not guidance. Running the file is covered in the local tools guide, including the warning that local is not automatically private.
What a team should file with the choice
Keep a one-page record: job, model file name, license link, hardware it ran on, whether tool use was required, and who tested it. When a new name shows up in a newsletter, update the record only after the card is read. Do not swap production to a new file because a chart moved. If the job involves untrusted text and actions like email or merges, add the controls in AI Agent Security for Safe Automation.
Questions worth asking before you buy
Ask what the license allows for your product, what hardware the file needs, and who will re-read the card when the name changes. A hosted API with a clear contract may be the better buy if nobody can maintain a local file.
For companies that need more than a subscription
If the company cannot accept the license, the model is not a candidate, even if the demo looked strong. A custom agent can call an approved model. It should not download a new one because a forum said it won.
If an off-the-shelf tool cannot meet your data, approval, or security rules, talk to AI Savr about a custom AI agent. Use the directory to compare live tools first, and vendors can submit a product.
Frequently asked questions
What is the best open source LLM right now?
The one whose license you can accept, whose size fits your hardware, and whose card matches the job. This page does not crown a benchmark winner.
Why are there no scores on this page?
A score this desk did not measure would be fiction. Model names also change faster than a roundup. The model card wins.
What should I read on the model card?
The license, the size, the exact file, and any note about tool use or limits. If those lines are missing, you are guessing.
Does a downloadable file mean I can ship it?
No. Download terms and a license for shipping a product can differ. Read the card before you build on it.
How do I judge tool use?
If the job calls functions or edits a repository, read whether the card documents that behavior. Then test it on your tasks.
How often do the names change?
Often enough that you should store the file name and the card link, not a nickname from a blog.