Open a laptop-friendly inference tool today, the kind that lists locally runnable models by download count, and the top of the chart is dense with one name: Qwen. Not a single popular release, a whole cluster of them, spanning sizes from barely more than half a billion parameters up past thirty billion. A year or two ago, if you’d guessed which lab would end up owning the “runs well on hardware you already own” category, Alibaba would not have been the first name most people wrote down. The path that got Qwen there is less a single breakthrough and more a pattern repeated often enough to become a habit.

Coverage before scale

Qwen2 and Qwen2.5, both released in 2024, set the pattern early. Rather than one flagship checkpoint, Alibaba shipped a range of sizes at once, most of them under the Apache 2.0 license, which let developers and companies use, modify, and redistribute the weights with almost none of the restrictions that come with more guarded “open” releases. Alongside the general-purpose models came dedicated variants: a Coder line tuned for programming tasks, a Math line tuned for quantitative reasoning. Instead of asking one model to be good at everything, Qwen asked several smaller models to each be good at one thing.

Then came QwQ-32B in November 2024, a comparatively compact model built specifically around reasoning rather than raw breadth. At 32 billion parameters it was small enough to run on serious but not exotic hardware, and it staked out reasoning as a capability worth isolating and tuning for directly, ahead of reasoning-focused releases becoming standard practice across the industry.

Qwen3 and the full spread

Qwen3, released in April 2025, is where the strategy became a full lineup rather than a handful of releases. Everything in it shipped under Apache 2.0: dense models running from 0.6B up to 32B parameters, plus mixture-of-experts versions for people who wanted more capability behind a similar active-compute budget. The whole family was trained on 36 trillion tokens and covers 119 languages, a range that matters as much for who gets to use these models as for how capable any single one of them is. Qwen3 also introduced a hybrid mode that lets a user toggle between “thinking,” where the model reasons step by step before answering, and “non-thinking,” where it answers directly for speed, along with native support for the Model Context Protocol (MCP), the emerging standard for connecting models to external tools and data.

Qwen3.5 and Qwen3.6, arriving through 2026, kept the same cadence but narrowed the focus toward agentic coding: models meant not just to write a function on request but to operate inside a coding workflow, calling tools, reading output, and iterating. Notably, Qwen’s actual flagship, the proprietary “Max” model, has stayed closed throughout all of this. The openness is a strategy applied everywhere except the very top.

Qwen won the local-model category with close to the opposite approach of the labs chasing frontier scale. It’s less like a single luxury flagship store and more like a hardware chain that stocks something for every shelf, from the budget aisle to the specialist counter, updated on a predictable schedule rather than around one blockbuster launch a year. That mostly-Apache-2.0 catalog, with the flagship carved out as the one closed exception, is a useful concrete case for the distinction drawn in Open Source vs Open Weight, and the sheer range of sizes on offer only pays off if you’re also matching a model’s size to the hardware that can actually run it, the subject of Hardware for Open Models.