Local AI for English: how to choose Gemma, Qwen or Ministral

For an English-speaking local AI assistant, shortlist a downloadable model that fits your computer, then test it on your actual writing and documents. Gemma 4 12B, Qwen3.5 9B and Ministral 3 8B are three candidates with published weights. Their developers, architectures and runtime requirements differ. None is a universal winner simply because it supports English.

Documentation checked on 13 October 2026 by the BestLLMFor editorial team. This guide is a source-based shortlist, not a newly conducted hardware benchmark.

Three models worth evaluating

The selection deliberately includes developers from different countries. “English-language AI” describes your task, not the nationality of its creator. These are specific releases, not a claim that they are the newest or best models available.

Start with the decision your assistant must make

A writing assistant should preserve meaning while improving clarity. A document assistant should stay within the supplied evidence. A coding assistant should produce code that passes meaningful checks. Combining these into one unexplained score can hide the failure that matters most to you.

Write down three recurring tasks before downloading anything. For each task, define an unacceptable mistake: inventing a delivery date, changing a quoted amount or returning prose when the application expects JSON. This gives you a usable comparison even without a large benchmark suite.

What “runs locally” actually means

Local inference means the model executes on your machine. It does not establish where every component of the surrounding application runs. Search, document indexing, extensions and account synchronisation may involve other services.

After installing the required files, test the intended workflow without an internet connection. Then inspect the application's settings and network behaviour before describing the whole workflow as private. A successful offline test establishes that the tested operation works offline; it does not prove what the application does when reconnected.

Choosing a size without guessing RAM requirements

Model size is not the same as required memory. As a rough calculation, 12 billion parameters stored at 4 bits occupy about 6 GB in decimal units for the raw weights alone. Runtime buffers, quantisation overhead and the context cache require additional memory.

Check the actual download size and supported architecture. A model name in a software catalogue does not establish that every quantisation, vision feature or long-context setting works on your device. Start with short text-only tasks, record peak memory and increase document length gradually.

An English test you can reuse

Run these original prompts on each candidate with the same output limit. Save the model version, quantisation, application version and responses.

1. British English editing: “Rewrite for a UK customer without changing any facts: Your order BX-17 shipped on 12 October. We expect delivery on 16 October, but cannot guarantee it.” Reject an answer that turns the estimate into a promise.
2. American English extraction: “Return JSON only, with order_id and total_usd: Order US-204 has a total of $1,249.90.” Check valid JSON, the exact identifier and the numeric value 1249.90.
3. Evidence boundary: “Use only this note. Who approved the purchase? Note: The purchasing team received the request on Monday.” A reliable response must say the approver is not stated.

Repeat each task three times. Track factual mistakes, formatting failures and total response time separately. More fluent prose should not compensate for invented information.

Frequently asked questions

Do I need a model developed in an English-speaking country?

No. You need satisfactory performance on your English tasks. Test regional spelling, vocabulary, dates and tone explicitly rather than assuming that a multilingual label guarantees them.

Are these models completely free?

The cited releases publish weights under Apache 2.0. Hardware, electricity, hosted inference or an application subscription may still cost money. Check the exact distribution and service you intend to use.

Can I compare advertised benchmark scores directly?

Only when the task, dataset, prompting and evaluation conditions match closely enough. Publisher scores can help identify candidates, but they do not replace a test of your workflow.

Continue with BestLLMFor's local AI selection tools and guides to match a candidate to your use case and computer.