Quick diagnosis
If a local model loads but invents facts, ignores instructions, repeats itself, or gives an incomplete answer, separate missing information from model capability and configuration. A larger context setting does not supply knowledge the model never received, and a fluent answer is not evidence that it is correct.
This guide suits local chat, writing, coding, and document work after the runner is functioning. A process that crashes or cannot load belongs in the memory or model-loading diagnostic.
First three checks
- Open a new chat with one clear task and a short, relevant source passage. Remove old instructions, web tools, and unrelated attachments for this check.
- Confirm the exact downloaded model and its chat template. An instruction model, a base model, and a coding model are different choices.
- Compare the answer against facts you can independently identify in the passage, including conditions, dates, numbers, and missing information.
A good result follows the requested format, preserves the supplied facts, and leaves unsupported facts unknown. Getting one example right is a useful diagnostic result, not proof of reliability for every task.
Give the model an answerable test
Use a harmless sample such as:
Source: The draft may be sent on Friday if both reviewers approve it. One reviewer has approved it. The meeting is at 10:30. No fee is stated.
Using only this source, state the meeting time, whether Friday delivery is confirmed, and the fee. Quote the words supporting each answer. Say “not stated” when the source gives no answer.
The expected facts are 10:30, delivery still conditional on both approvals, and no stated fee. An answer claiming that delivery is definitely Friday or inventing a fee reveals a meaningful error even if its tone is polished.
For writing, ask for a rewrite preserving the same conditions and check it line by line. For coding, use a scratch project with a small function and known input/output cases, then run those checks yourself. See local writing models and local coding models for practical comparison tasks.
Find out whether the source reached the model
In document chat, the app may supply the whole text of a short file or retrieve a few passages from a larger one. An attached file icon does not prove that all its contents were included. Ask about a known sentence and open the citation to inspect the original passage.
A scanned PDF may have no usable extracted text. Tables, footnotes, columns, and page numbers can also become distorted. Check the extracted text before changing the model. If the answer depends on two sections, supply both with their headings rather than removing the cross-reference to make the task easier.
The document guide explains retrieval and OCR limits. The token counter can help plan how much text you are supplying, but another model's tokenizer and the app's added instructions can change the actual count.
Check model and conversation settings
Use a chat or instruction-tuned model with the publisher's intended template. The template formats user messages and system instructions; a mismatch can cause strange formatting, repetition, or failure to stop. If you imported a custom download, return to a supported download with its supplied template before making experimental edits.
Begin in a fresh chat with enough context for the task and answer. Earlier messages and tool responses share that space. If the answer cuts off, check the output-token limit and stop conditions as well as context. In Ollama, num_ctx controls context and num_predict controls generated output; changing one does not change the other.
A lower temperature can make responses less varied, but cannot establish truth or add missing knowledge. Use the model publisher's recommended settings when available. Qwen3's /no_think instruction is specific to compatible Qwen3 thinking templates; it is not a universal repair command for other models.
Change one thing and compare the same task
Keep the source, prompt, and answer requirements fixed. Try a supported instruction model, then a larger or less aggressively compressed version if the hardware calculation allows it. Check whether it preserves the same dates, exceptions, and absent facts. Avoid comparing a tiny model's difficult prompt against a larger model's easier prompt.
If both fail, revisit the source extraction and instructions. If one succeeds but requires heavy swapping or long waits, it may be impractical for your computer even when its answer is better. Keep both correctness and usable response time in the decision.
For current events, product versions, or external authorities, provide the current source or consult it directly. A model running offline cannot independently fetch a newly changed fact. For legal, financial, or other consequential work, verify each material statement against the original authority before relying on it.