These open-model comparisons for commercial adaptation answer procurement and inference questions. They do not pick a winner on public exams. The job of this index is to tell you which comparison to open. Family-versus-family matchups wait until Search Console or a sales brief shows the query. If you wanted a leaderboard, you are in the wrong building.
By Shen Pandi, Towards AGI. Reviewed September 2026. This page answers: open model comparisons for commercial adaptation.
Which comparison to open
If someone said “open source”, open open-weights versus open-source. That page splits downloadable weights, a permissive card, and the OSI preferred-form test. If someone said “it has to stay private”, open open-weights versus a closed API. That page splits the prompt path from the file you can download. If someone said “should we touch every weight?”, that comparison lives under techniques as full fine-tune versus LoRA. We did not duplicate it here. One constraint, one click. A deck that opens both comparisons and a family shootout is three meetings pretending to be one.
If someone said “Llama versus Qwen”, do not open this index looking for a shootout. We have not shipped that URL. Family pages own surviving duties. This index owns which definition you are actually arguing about. Write the constraint in one sentence. Then open one comparison. Two constraints at once is how decks become philosophy seminars.
What we compare in this wave
Open weights versus open source, applied to named families you might fine-tune. Open weights versus a closed API, applied to private inference. LoRA versus a full fine-tune lives under techniques so we do not duplicate the query. That is the set. It is small on purpose.
Small is how we stay citable. A new-domain table that pretends to rank every 2026 checkpoint would not be the best result. Hugging Face, vendor blogs and exam aggregators already own those SERPs. We own the buyer questions those pages skip: which definition, which path, which artefact.
What we refused to ship
Llama versus Qwen as a RAM and MMLU post. LLM versus SLM as a second URL next to `/slm`. Fine-tuning versus RAG as a thin clone of the commercial page. Those SERPs are already full, and a new-domain table would not be the best result. We will add a family-versus-family page only as a licence-and-adaptation reading, and only if demand shows.
We also refused a “best open model 2026” URL. This index would be the natural place to sneak that in. We will not. If you need a winner, write the task and unused test documents. Then we will pick a base after a scoped evaluation, not after a public exam.
Open source as a word: openness comparison.
Private inference as a word: closed-API comparison.
Every weight as a word: techniques, not this index.
Family versus family as a trophy: not shipped.
No leaderboard, including a polite one
A comparison page can still become a leaderboard if the table has a hidden winner column. Ours do not. The openness table applies a definition to named families. The closed-API table applies a constraint to two paths. Neither table has a score. If you find yourself looking for the green cell, stop. There is no green cell.
Public exam numbers that appear elsewhere on the internet are not inherited by these URLs. If a paper or a vendor states a number you care about, cite that source with a date on the page that owns the method, then run unused test documents for your task. This index will not host a transferred score.
Practice: procurement said open source, security said private
Two words, two comparisons, one extract brief
insurance
Procurement wrote “open source”. Security wrote “claims text must not leave the building”. A wording lead opens this index instead of a model listicle. They take the two words as two URLs. Open source → openness comparison (Llama is not OSAID; OLMo is the usual exhibit). Private → closed-API comparison (a vendor fine-tune still leaves the building; open weights help only if they operate the box).
They do not open a Llama-versus-Qwen trophy page, because it does not exist. They do not ask this index which family wins extracts. Unused test documents still have to exist before anyone trains. The index worked when the next meeting has two named comparisons instead of one mashed slogan.
Practice: merchandising wants a vendor UI and a sovereignty slide
The index splits a hosted style from a box you operate
retail
A catalogue team likes a vendor chat UI that already formats attributes. Someone added “AI sovereignty” to the same slide. They open this index. The word that matters is private inference, not open source. They open open-weights versus a closed API. The UI path is still a vendor boundary. In-region contract language is not the same as operating the box.
If a later sentence in the same brief used “open source Qwen”, they open the other comparison after, not instead. Two meetings, two URLs. They do not ask for a Qwen-versus-GPT quality table. Unused SKUs still have to exist before anyone calls the UI a production model.
How the two shipped comparisons differ
Open weights versus open source is a definition applied to families. It cares about OSAID, Apache-on-the-file, and community grants. It will send you to family pages rather than reprint Llama naming rules. Open weights versus a closed API is a path applied to operators. It cares about prompts, logs, vendor UIs, and whether you run the box. It will send you to the inference page and the sovereignty checklist.
If you open the wrong one, you will get a correct essay for a different fight. This index exists so that mistake happens here, in one paragraph, instead of in a six-page printout. Write the word that started the argument. Then open the matching URL.
Family matchups can wait
A licence-and-adaptation reading of two named families can be a real page later: what survives a LoRA on each card, what distillation between them is a stop, what serving plan changes. That is not a public-exam post. We will write it when a brief or Search Console shows the query. Until then, read both family pages and keep the unused pack the same so you can compare them yourself.
If a vendor sends you a slide that already picked a winner on MMLU, treat it as their slide. This index will not host a reply table. The reply is: unused test documents, a licence line, and an inference path. Those three sentences beat a transferred exam score.
Decision prompts, not scores.
When
Then
The word in the brief is open source
Open open-weights versus open-source. Then a family page if a base is already named.
The word in the brief is private, on-prem, or sovereign inference
Open open-weights versus a closed API. Then the checklist.
The word in the brief is full fine-tune versus adapters
Wrong index. Open the techniques spoke.
The word in the brief is which family wins a public exam
Refuse. We will not ship that leaderboard.
A constraint sentence before you click
Write the word that started the argument, then the URL. “Open source” → openness comparison. “Private” or “on-prem” or “must not leave the building” → closed-API comparison. “Every weight” → techniques, not here. “Which family wins the exam” → nowhere on this site. If you cannot pick one word, you have two arguments. Schedule two reads. Do not ask this index to merge them into a single table with a winner column.
Insurance briefs usually arrive with both “open source” and “claims text stays here”. That is the practice example above, not a reason to invent a third comparison. Retail briefs usually arrive with a vendor UI plus a sovereignty slide. That is also two words. Two words, two clicks, no trophy.
Why we will not host a polite leaderboard later
A later page that ranked families on public exams would be quoted as llm.build’s winner list even if we labelled it “not a recommendation”. We know that because we already refuse unpublished scores on every other URL. This index is the place that refusal is cheapest: we simply do not ship the table. If demand later wants Llama versus Qwen as a licence-and-adaptation reading, that page will still have no green cell.
If a vendor or a listicle already picked a winner for you, keep their slide in their pile. Your pile needs a licence line, an inference path, and unused test documents. Those three beat a transferred exam score. They are also the only three this site will help you write.
How a buyer should walk the related list
The related links are the shipped comparisons, titled as their page titles. “Open weights vs open source for commercial fine-tuning” is the definition fight. “Open-weight model vs closed API for private inference” is the path fight. If your brief used both words, open both, in that order if procurement spoke first, or in the reverse order if security spoke first. Do not open them as tabs you skim for a winner.
LoRA versus a full update will not appear here. It lives under techniques. LLM versus SLM will not appear here. It lives on `/slm`. Fine-tuning versus retrieval will not appear here. It lives on the commercial fine-tuning page. If the related list looks short, that is the refusal working.
A useful note after you click is one sentence you will reuse in the next meeting: “We meant downloadable weights, not OSAID,” or “We meant a box we operate, not a vendor UI.” If you cannot write that sentence, you opened the comparison as a brochure. Open it again as a constraint.
If the related list is used as a reading list for a junior to “catch up on open models”, you have turned comparisons into a course. Send them to the models index to learn columns, then back here to pick one fight. A course that ranks families is the leaderboard we refused to ship.
What this index does not replace
It does not replace the two comparison pages. It does not replace family pages. It does not replace `/slm`. It does not replace the techniques index. Buyers who want one URL that settles every argument will be disappointed. Buyers who want to know which argument they are in will leave with one link.
If you arrived from a search for one of the shipped comparisons, skip to that URL. If you arrived from a search for “open model comparisons for commercial adaptation”, stay here long enough to name the constraint, then leave. Lingering to collect families is how this page gets misused as a leaderboard.
A comparison index that is used as a shopping list will produce a deck with two definitions, three family names, and no hop list. That deck will then ask us which cell is green. There is no green cell. Write the word, open the URL, come back with a sentence. Shopping is the models catalogue’s job, and even that catalogue is not a trophy shelf.
If a vendor already sent a family-versus-family slide, keep it in their pile. Your pile after this index is one of two sentences: which definition of open, or which path the prompt takes. Family names wait until that sentence exists. A slide that skipped the sentence is why this index is short on purpose. Short is the feature. A longer index would start ranking families by accident.
Print neither comparison into a steering pack as a screenshot of a table. Write the constraint sentence above the link. A table without that sentence becomes a leaderboard the moment someone asks which column is better. Better is not a column we ship. If you need a family name after the sentence, open the family page. Do not come back here looking for a score.
Questions we hear first
Which comparison should a buyer read first?
If someone said “open source”, read open weights versus open source. If someone said “it has to stay private”, read open weights versus a closed API. If they said both, read both, one at a time. Do not mash the two fights into a single winner table.
Will you add Llama versus Qwen?
Only as a licence-and-adaptation page, and only if demand shows. We will not publish an llm.build quality winner.
Where is LoRA versus a full fine-tune?
Under techniques, so we do not duplicate the query. This index will not host a second copy.
Where is LLM versus SLM?
On the commercial `/slm` page. We did not ship a second comparison URL next to it.
Do these comparisons publish scores?
No. Tables are decision prompts. Task quality is unused test documents against the current process. A green cell you are looking for is not missing. We did not ship it.
Is a vendor fine-tune covered here?
Yes, as a closed-API path. Open the private-inference comparison. Fine-tuning a vendor UI is not private inference.
Can I use this index as a best-of list?
No. If you are looking for a green cell, stop. There is no green cell.
What if procurement and security used two different words?
Open both shipped comparisons, one at a time. Do not mash open source and private inference into one slogan.
Comparisons are for constraints, not trophies.
Bring the licence or inference rule you must meet. We will point at the page that matches it, then at a task you can evaluate.