The Qwen Family Explained: The Models You Can Own

Most of the open-weight field in 2026 is open in the licensing sense and closed in the practical sense. The weights are downloadable and the deployment needs a cluster. Qwen is the family that consistently ships models you can actually put on hardware you own.

That is a different value proposition from topping a leaderboard, and it is worth evaluating on its own terms rather than against models that were never going to run on your machines.

Dense is the defining choice

The important structural fact about Qwen 3.6 27B, released 20 April 2026, is that it is dense. Every parameter runs on every token, so memory and compute scale together and both are predictable.

Contrast that with a mixture-of-experts model, which holds a very large parameter store and routes each token through a small subset. That design is cheap per token and expensive to hold, because every expert must be resident somewhere even when most are idle for a given token.

The consequence is direct. A dense 27B fits on a single accelerator once quantised. A trillion-parameter mixture-of-experts does not fit on any single accelerator regardless of quantisation. Dense against mixture-of-experts covers why labs keep building the ones most teams cannot run.

The capability gap is smaller than the size gap

Qwen 3.6 27B reports 77.2 percent on SWE-bench Verified. DeepSeek V4 Pro — a 1.6T mixture-of-experts activating roughly 49B per token — reports 80.6 percent on the same benchmark.

That is a gap of 3.4 points between a model you can run on one card and a model that requires a multi-node deployment with fast interconnect. Stated as a ratio of parameters it sounds enormous; stated as a ratio of capability on this benchmark it is close to a rounding difference.

Be careful not to overread it in the other direction either. Benchmark gaps of a few points frequently sit inside the noise introduced by the evaluation scaffold, so the honest summary is that these two models are close on this measure and very far apart on deployment cost. What actually fits on one GPU works through the memory arithmetic.

Reading a family with many tiers

Qwen ships across a range of sizes, and the range is the point. A family that spans from models sized for a laptop to models sized for a cluster lets you develop against a small tier and deploy against a larger one without changing your prompts, your tokeniser assumptions or your API shape.

That consistency has practical value that never appears in a comparison table. Prompt behaviour transfers reasonably well across tiers of the same family, so an evaluation harness written against one size mostly keeps working against another.

A caution on the preview tiers. Qwen 3.6 Max is in preview at the time of writing, which means published figures are provisional and worth treating as such. Do not plan capacity or make a commitment on a preview specification — wait for the general release or measure it yourself. The Qwen 3.6 guide covers the generally available tier.

Licensing is mostly permissive

Qwen ships permissive terms on most tiers, which puts the family in a comfortable position for commercial use without the conditions that some other families attach.

The qualifier "most tiers" is doing real work in that sentence, and it is the reason to check the specific model card rather than assuming the family. Licences vary by tier and by generation, and the tier you evaluated may not be the tier you deploy.

Compare against Moonshot's Kimi models, which lead the open-weight field on agentic benchmarks and ship under Moonshot's own terms including a revenue threshold on K3 for offering the model as a service. If you are building a product around serving a model, the licence is a first-order concern, not a footnote. The licence comparison covers who ships what.

Quantisation is how it gets onto your card

A 27B model at 16-bit precision is roughly 54GB of weights before any cache or overhead. At 8-bit that halves; at 4-bit it halves again. That arithmetic is what turns "needs two accelerators" into "needs one", which is the only threshold that changes your architecture.

The cost is uneven. Quantisation tends to preserve conversational fluency well while eroding precise multi-step reasoning first, which is exactly the capability coding work depends on. A quantised build can pass casual testing and then underperform on the tasks you deployed it for.

The discipline is to evaluate the exact build you will ship, on your own tasks, rather than reading the benchmark for the full-precision original. Quantization explained covers which formats degrade gracefully.

When Qwen is the right answer

Choose it when data cannot leave your network, when your volume is high enough that per-token pricing has become the dominant cost, or when you need availability that does not depend on a vendor's uptime. Those are the three durable reasons to self-host, and Qwen is the family that makes them practical.

Choose it also for latency-sensitive work you can put close to your users, and for anything you intend to fine-tune — dense models have far more mature tuning tooling and none of the routing-collapse failure modes that make mixture-of-experts tuning awkward.

Look elsewhere when you need the strongest available agentic performance and you are happy on an API, or when your workload genuinely needs a million-token window. And re-test periodically: training quality has been improving faster than scale, so a conclusion you reached about small models two years ago may no longer hold. Choosing a model for self-hosting covers the wider shortlist.