The Qwen Open-Weights Family, Explained for People Who Just Want It to Work

Alibaba’s Qwen team releases models at a pace that makes documentation obsolete before it is published. Several a month, spanning sizes from under a billion parameters to hundreds of billions, in dense and mixture-of-experts variants, with separate instruct, thinking, coding, and multimodal lines.

It is also, by download volume, the most used open-weights family in the world. If you run a local model, there is a good chance it is a Qwen. Here is how to navigate the family without reading every release note.

The Naming Decodes Cleanly Once You Know The Pattern

A typical model name carries four pieces of information: the generation, the parameter count, the variant, and the quantization.

Generation is the number after Qwen. The Qwen3 line has been the workhorse through 2026, with 3.5 and later refreshes extending it. Each generation is a genuine retrain, not a tune.

Parameter count is the number followed by B. For mixture-of-experts models you will see two numbers, as in 35B-A3B: 35 billion total parameters, roughly 3 billion active per token. The active count drives speed; the total count drives memory. This distinction confuses people constantly, and it matters: an MoE model needs memory for all its parameters even though it only computes with a slice.

Variant tells you the training objective. Instruct models follow instructions and are what you want for rewriting. Base models are raw completion engines and are not what you want. Thinking or reasoning variants emit a deliberation phase before answering, which helps on hard problems and gets in the way on simple transformations.

Quantization appears as Q4_K_M, Q8_0, or similar in GGUF filenames. Q4_K_M is the widely used four-bit compromise: roughly a quarter of the memory of the full-precision model, with degradation small enough to be acceptable for most work.

So Qwen3-4B-Instruct in Q4_K_M is a fourth-generation four-billion-parameter instruction-following model, four-bit quantized, at roughly 2.4GB on disk.

The Sizes That Actually Matter For Writing

Most of the Qwen catalog is irrelevant to professional writing, and that is fine. Three tiers are worth knowing.

Around 1.7B. The practical floor for reliable instruction-following. Roughly 1.1GB on disk quantized, about 1.7GB of RAM in use, fast on CPU alone. Wrivio ships Qwen3 1.7B as its Standard local model precisely because it clears the instruction-following bar on a machine with no GPU.

Around 4B. The step up when you have memory to spare. About 2.4GB on disk, roughly 3GB in use. Noticeably better at holding several constraints simultaneously, which matters when your rewrite instruction specifies register, length, and fact preservation at once. This is Wrivio’s Best tier.

Around 30B, MoE. Genuinely strong, needs a real GPU. A 30-something-billion-parameter MoE with 3B active runs well on 24GB of VRAM with int4 quantization and is competitive with much larger models. Worth it if you have the hardware and want local output that is hard to distinguish from cloud.

Everything above that is server territory. The current model cards live on Hugging Face and the code and release notes on the QwenLM GitHub organization.

The Thinking-Mode Detail That Trips People Up

Some Qwen3 models are hybrid: they can operate in a deliberating mode or a direct mode. If you do not explicitly select direct mode, the model may spend tokens reasoning about a task that needs no reasoning, which for rewriting means slower output and occasional visible deliberation leaking into your email.

The fix is a template detail rather than a setting: the chat template prefills an empty thinking block to signal non-thinking mode. Tools that support these models properly handle it for you. Wrivio does this for hybrid models, which is why a local rewrite returns prose rather than a monologue about prose.

If you are wiring a model up yourself, this is the single most common cause of “why is my local model so slow and chatty,” and it is worth checking before blaming the model. We covered the tradeoff in thinking models versus instant models for rewriting.

Licensing: Generally Clean, Still Worth Checking

The Qwen3 open releases have generally shipped under Apache 2.0, which permits commercial use with attribution and carries no user-count restriction. That is a meaningful improvement over the family’s earlier history: some Qwen 2.5 models were published under a research license that forbade commercial use.

Wrivio hit exactly this. An earlier build shipped a Qwen 2.5 model under the non-commercial research license, and the fix was moving to Apache 2.0 Qwen3 weights. The lesson generalizes: license terms vary within a family and across generations, so check the model card for the specific version rather than trusting the family name. We wrote up the categories in open-weights model licenses explained.

Multilingual Strength Is The Underrated Feature

Qwen’s training emphasis on Chinese and English, plus broad multilingual coverage, makes the small models unusually good for non-English professional writing at their size. If you write work email in a second language, this is where the family earns its place: a 4B model that handles your register correctly in German or Spanish is more useful than a frontier model you have to open a browser for.

Test it rather than assuming. Language coverage at small sizes is uneven, and the only benchmark that matters is your own text. There is more on this in open-weights models for non-English writing.

Common Questions

Which Qwen should I start with?

If you have 8GB of RAM and no GPU, start at 1.7B. If you have 16GB or more, start at 4B. Do not start at 30B unless you have a 24GB GPU, because CPU inference at that size is slow enough to change your behavior.

Instruct or thinking?

Instruct for rewriting, summarizing, and reformatting. Thinking variants for problems with a right answer that requires several steps. Rewriting is not that kind of problem.

What does A3B mean?

Three billion active parameters per token in a mixture-of-experts model. It predicts speed, not memory. You still need enough RAM or VRAM for the full parameter count.

Is Qwen safe to use for confidential work?

The weights running on your own machine make no network calls, which is the property that matters. Using a hosted Qwen API is a different situation with the usual cloud retention and jurisdiction questions.

Download Wrivio for Windows to run an Apache 2.0 Qwen3 model in-process, with the thinking-mode and template details already handled.