Researched October 27, 2026. These recommendations match ListingSail Desktop's current local JSON adapter. Published model specifications establish capability and availability; they are not ListingSail benchmarks. No model was downloaded or called for this guide.
Start with qwen3:4b-instruct-2507-q4_K_M if your computer has enough memory. It is a compact, non-thinking instruct model with a verified Ollama tag. Use the cloud AI model guide to compare other providers, the documentation center for campaigns, or Connect your AI for connection choices. A downloadable cloud AI guide is also available.
Verified model choices
Choose the complete tag, including the size and quantization. A family name or latest alias may resolve to a different model. The download sizes below were shown by the official registry on the research date; they are not RAM requirements.
Larger multilingual alternative when memory permits
The exact Qwen3 Instruct 2507 model supports only non-thinking mode. This avoids making a separate reasoning trace part of the normal response workflow. Qwen2.5's authors explicitly describe improvements in structured data and JSON generation. These are reasons to try the models, not promises of factual accuracy. Sources: Qwen3 model card, Qwen2.5 model card.
Hardware estimates
These are planning estimates, not vendor minimums or measured results. Allow room for Windows, the browser, model buffers and the context cache. The table assumes one model and roughly 8K–16K context. Longer context and concurrent workloads need more memory.
Model size / example
Suggested system RAM
Optional GPU VRAM target
Practical expectation
1.5B Qwen2.5
8–12 GB
GPU optional; about 4 GB if used
Prefer short requests; full website extraction may be too demanding
4B Qwen3 Instruct
16 GB
About 8 GB
Sensible first configuration for this guide
7B Qwen2.5
16–24 GB
About 8–12 GB
More memory and processing than the compact option
14B Qwen2.5
32 GB
About 16 GB
Check actual latency before selecting it for campaigns
4B Gemma 3
16 GB
About 8 GB
Alternative to compare with the compact Qwen option
12B Gemma 3
32 GB
About 16 GB
More room for model and context overhead
CPU inference can work when system memory is sufficient, but a model fitting in RAM does not mean it will finish within ListingSail's timeout. Cold loading, prompt length, CPU/GPU placement and other running models affect latency. Larger is not automatically better for this workflow. Use ollama ps to inspect CPU/GPU placement. Sources: Ollama memory and placement FAQ, context and memory.
English and Bulgarian
Qwen3's official language list explicitly includes Bulgarian, and the Instruct 2507 card describes multilingual improvements. That makes it a reasonable starting candidate for both English and Bulgarian source text; it is not a task-specific quality guarantee. Sources: Qwen3 language coverage, exact Instruct model card.
Qwen2.5 documents more than 29 languages and JSON capabilities, but its cited language examples do not individually establish Bulgarian quality. Gemma 3 documents over 140 languages; evaluate Bulgarian names, addresses, descriptions and quoted evidence yourself. Sources: Qwen2.5 card, Google Gemma 3 card.
ListingSail asks for website copy in the source language. English product navigation does not mean that Bulgarian website descriptions should be silently translated. Review facts in either language before granting submission permissions.
Choose a model, then explicitly download it yourself. ListingSail does not download or install models. For the recommended option, run these commands in your own terminal:
ollama --version
ollama pull qwen3:4b-instruct-2507-q4_K_M
ollama ls
Open the current installed ListingSail from the Windows Start menu. Ollama support is available from Desktop 0.1.3; an old portable copy can show different settings.
Open AI & connections. Set Provider to ollama and Model to qwen3:4b-instruct-2507-q4_K_M, or the exact name of another already installed suitable model.
Choose Save settings, then Test AI connection. This performs a small inference test. It does not download a model, submit a website or establish extraction quality.
The current adapter connects only to http://127.0.0.1:11434. It accepts installed local GGUF weights and rejects cloud-labelled or remote-model metadata. An arbitrary endpoint, a model file path or a :cloud tag is not supported. Local inference does not require an AI-vendor key in ListingSail, but the ListingSail license and its own checks still apply.
You can inspect installed-model metadata without generating text:
A model's advertised maximum context is different from the context actually configured in Ollama. Current Ollama context documentation gives a 4K default below 24 GiB VRAM. That can be insufficient for a longer ListingSail website import: the app can supply up to four pages with approximately 40 KB of public text in total, plus instructions and the output schema. The current adapter does not set num_ctx. Source: Ollama context documentation.
For a model-specific configuration, first download the recommended base model. Then save this text as a file named Modelfile:
FROM qwen3:4b-instruct-2507-q4_K_M
PARAMETER num_ctx 16384
Create a local alias using your terminal:
ollama create listingsail-qwen3 -f .\Modelfile
ollama ls
Select listingsail-qwen3 as the model in ListingSail. The alias references the installed base model and applies its context setting. It does not change ListingSail's fixed endpoint or the base model's weights. Source: Ollama Modelfile reference.
16,384 tokens is a starting point, not a guarantee that every four-page request fits. Long multilingual text may need 32,768 if the model and available memory permit it. Leave space for instructions and generated output. More context increases memory and processing costs; do not choose a huge context merely because the model advertises it. After a request, ollama ps or the local /api/ps metadata can help verify the loaded configuration. Sources: context sizing, running-model metadata.
Test connection and actual quality separately
The connection test checks that the selected provider can return a small usable response. Then evaluate a real import before using it in a campaign:
Use a public website whose facts you can verify. Keep publication and registration permissions off during evaluation.
Compare the suggested name, description, contacts, pricing and source quotes with the actual pages. Missing facts should stay missing.
Repeat with Bulgarian source text if that is part of your work. Check Cyrillic names and addresses without assuming that multilingual support guarantees correctness.
Review a prepared directory form before sending it. A correct field mapping is different from an accepted submission or a live listing.
ListingSail's daily AI-call budget still applies to local requests. A failed or uncertain call can reserve an app call even though it does not create a model-vendor charge. Local model use consumes your own hardware resources. See the main documentation for permissions and campaign states.
Current adapter limits
The current local adapter sends stream: false, a JSON schema in format, an empty tools list and temperature 0. It allows up to 4,096 generated tokens and approximately 120 seconds for an operation. The output is validated; generated code is not executed. Ollama documents JSON-schema output and low-temperature generation in its structured-output guide.
The adapter does not explicitly set thinking controls. A generic thinking model can spend its time or token allowance before producing useful JSON. The exact recommended Qwen3 Instruct variant avoids that default; do not assume that qwen3 or qwen3.5 is interchangeable with it. Qwen3.5's official card says it thinks by default. Sources: Ollama thinking controls, Qwen3.5 model card.
This workflow uses text helpers for profile extraction, field mapping and bounded choices. Vision, tool-use or agent labels on a model page do not add new ListingSail capabilities. Local inference also does not make the whole product offline: license validation and public website/directory work still need their respective connections.
Troubleshooting
Ollama is missing from the provider selector: check the app version and open the installed application instead of an older extracted ZIP. Current connection choices are explained at Connect your AI.
Connection refused: start the Ollama application and check local metadata at 127.0.0.1:11434. ListingSail does not start the service or accept a replacement URL.
Model not confirmed as local: compare the exact tag with ollama ls and the /api/tags metadata. A cloud model, missing model or incompatible format will be rejected.
Timeout or exhausted memory: try a smaller model or a shorter request. Check CPU/GPU placement and context usage. A large model that loads successfully can still be too slow for this operation.
Valid connection, poor profile results: inspect source evidence and configured context. A short connection test can pass while longer input is truncated or produces omissions. Compare another suitable model rather than assuming a larger one is always correct.
Repeated failures: inspect the cause before retrying. ListingSail does not automatically retry or switch providers; failed attempts can still use the daily app allowance.
Recommendations and registry sizes were checked on October 27, 2026. Model availability, tags and vendor behavior can change. No speed, accuracy or Bulgarian extraction benchmark is claimed for ListingSail.