Principles
- Conditional, testable claims. "Private", "offline", "cheaper" and "compliant" depend on configuration and workload. We say what has to be true for a claim to hold.
- Customer acceptance criteria outrank our preferences, including any wish to recommend tools we build.
- A recommendation to stay hosted, or do nothing, is a valid result.
Evidence labels
Every performance or cost figure we publish carries one of these labels:
| Label | Meaning |
|---|---|
| Measured by us | We ran it, on recorded hardware, model, runtime and quantisation versions, with stated input/output lengths and concurrency. |
| Reproduced externally | An independent party has reproduced our result, or we have reproduced theirs, under recorded conditions. |
| Vendor-reported | Published by the vendor or project. We have not verified it. |
| Estimated | Derived from a formula or rule of thumb, such as memory bandwidth ÷ model size. Useful for orientation, never a substitute for testing your workload. |
For measured results we record: model and runtime versions, hardware, quantisation, context and output lengths, concurrency, time to first token, end-to-end latency, resource use and task-quality results. For security-sensitive workflows we also test permissions and disclosure, not just answer quality.
Our own products
Our team (Cognisoc) builds open-source local AI tools, including Mullama, Llamafu and ZigLLM. Wherever we cover or recommend one of them, the page says so visibly. They are listed and compared on the same terms as other tools, and are never a required part of a client engagement.
Sponsorship and listings
- Sponsored content is labelled as sponsored at the top of the page.
- Rankings, "best of" picks, comparison verdicts and benchmark conclusions are never for sale.
- Tool submissions are reviewed editorially. Submitting does not guarantee coverage or position.
- We do not currently use affiliate links. If that changes, affected links will be marked.
Services and referral fees
Neul Labs Limited, which owns and operates local-llm.net, sells assessment, deployment and support services through local-llm.net. Assessments are paid, fixed-fee deliverables, so a recommendation against migration is still complete, paid work.
If a hardware, hosting or implementation supplier pays us a referral fee, we tell the customer in writing before they buy, and the customer can always buy directly. Referral fees never decide a recommendation.
Client data and case studies
Client documents and test material stay in the environment the client approves. Only results the client authorises leave it. Case studies are published only with permission, anonymised unless the client asks otherwise.
What we publish, and what we don't
We publish a page when there is a distinct question and enough evidence to answer it usefully. We don't mass-produce near-identical pages for every model × GPU × industry × country combination. Where we have such pages from the past, we label their figures as estimates or keep them out of search indexes.
Corrections
If you find an error, especially in a number, email contact@local-llm.net. We fix material errors promptly and update the page's "Updated" date. In October 2026 we corrected cost and payback calculations across several guides that had overstated savings, some by mixing up weeks and months.