Qwen Models Explained
Qwen is not one model. It is Alibaba’s whole family of AI models — a family that, as of 16 November 2026, has a brand-new flagship, Qwen4-Max, which you cannot download at all. The flagship is available only through Alibaba Cloud’s API. The models you can actually download and run on your own machine sit one generation behind, in the Qwen3.8 line, led by the practical default Qwen3.8-27B. That split — a closed flagship on top, an open line underneath — is the single most important thing to understand about Qwen in November 2026, and it is the spine of this guide. Checked 17 November 2026 against Qwen’s own pages, the Qwen3 Technical Report, and Alibaba Cloud’s Model Studio pages as reported by multiple secondary reads.
There is a second confusion to clear up before anything else. “Qwen” is four things at once: the name of the model family, the name of the flagship model inside it, the name of the free chat app at chat.qwen.ai, and Alibaba’s AI brand in general. On top of that, the version numbers do not mean what version numbers usually mean. Qwen3.5, Qwen3.6, Qwen3.7 and Qwen3.8 are sub-generations, not a simple march upward — and Qwen3.7 was never released as open weights at all, so if you go looking for a Qwen3.7 download, you will not find one. This guide maps the whole family in plain language: every generation, every specialist branch, what is open and what is closed, what each model costs to use, and which one you should actually pick.
Two corrections belong at the top, because most explainers written even a few weeks ago get them wrong. First, Qwen4 exists now. It is not a rumour or a roadmap item: Qwen4-Max was released on 16 November 2026, the day before this guide was checked. Second, QwQ — the reasoning model that made Qwen famous in early 2025 — is history. Its last release was in March 2025, there is no successor, and its reasoning ability was folded into Qwen3’s switchable “thinking mode”. Search interest in QwQ is still high, but the answer readers need in November 2026 is not a QwQ how-to; it is that thinking mode is what QwQ became. Prices in this guide are date-stamped “checked 17 November 2026”, because Alibaba can change them, and the pricing page is the authority if the figures here have moved.

What Qwen Actually Is — One Name, Four Different Things
<img src="

” alt=”A person looking puzzled at a wall of many similar-looking blank boxes, a visual metaphor for the confusion caused by one name meaning several things. No readable text.”>
Qwen — Chinese name 通义千问, Tongyi Qianwen — is the AI model family built by Alibaba Cloud and the Qwen team. It began in 2023 as a single line of large language models and has grown into something much wider: general chat models, reasoning models, coding models, vision models, image models, translation models, embedding models, a safety classifier, a speech model, a chat app, and a terminal coding agent. When someone says “Qwen”, they could mean any of four things, and the first step to understanding the family is knowing which one is meant.
The first meaning is the family itself: every model Alibaba’s Qwen team has released, from the original Qwen in 2023 through Qwen2.5, Qwen3, Qwen3.5, Qwen3.6, Qwen3.7, Qwen3.8 and now Qwen4. The second meaning is the flagship model of the moment — right now, Qwen4-Max. The third is the product most people actually touch: Qwen Chat, the free chat app at chat.qwen.ai, and the Qwen app on phone and desktop, which run the latest models behind a simple chat box. The fourth is the brand: “Qwen” as Alibaba’s name for its AI effort as a whole, the way “Gemini” is Google’s and “Claude” is Anthropic’s. None of these meanings is wrong. The trouble starts when a guide uses one meaning and the reader hears another — when “Qwen is open source” is taken to mean the flagship is downloadable, for example, when it is the older generation that is open and the flagship that is not.
The practical consequence is that almost every question about Qwen needs a second question: which Qwen? “Can I run Qwen locally?” has a different answer for Qwen3.8-27B (yes, on a single good graphics card) than for Qwen4-Max (no, it is API-only, with no downloadable weights). “Is Qwen free?” has a different answer for Qwen Chat (free to try) than for the API (paid per token) than for the open weights (free to download, but you pay for the hardware to run them). “Is Qwen open source?” is the trickiest of all, because the honest answer is “mostly open weights, with exceptions, and open weights are not the same as open source” — a distinction the licences section of this guide takes up in full. Keep the four meanings apart and the rest of the family tree falls into place.
One more piece of framing helps. Qwen is the most prolific open-weight model family in the world by a wide margin. The Qwen organisation on Hugging Face lists 469 models in 37 collections, read directly on 17 November 2026. No other model family comes close to that breadth: general models in sizes from under a billion parameters to trillions, plus dedicated lines for code, vision, speech, images, translation, search and safety. That breadth is Qwen’s greatest strength and also its greatest source of confusion, because a family this large cannot be summarised by naming one model. The rest of this guide is the map.
The Qwen Family Tree at a Glance

The general line of Qwen has moved through eight generations in three years. The table below is the whole story in one place: generation, date, the headline, and whether you can download it or only call it through an API. The specialist branches — coding, vision, image, and the rest — sit alongside this spine and are covered in their own sections below.
| Generation | Date | Headline | Open or API-only |
|---|---|---|---|
| Qwen | 2023 | The original line; Alibaba’s first large language models | Open weights |
| Qwen2 / Qwen2.5 | Sep 2024 | The line that made Qwen a household name among developers; the base for QwQ and Qwen2.5-Coder | Open weights |
| Qwen3 | 28 Apr 2025 | Thinking and non-thinking modes in one model; the release that changed everything | Open weights, Apache 2.0 |
| Qwen3.5 | Feb–Mar 2026 | Natively multimodal; sizes from 0.8B to 397B | Open weights, Apache 2.0 |
| Qwen3.6 | Apr 2026 | Refinement generation; 35B-A3B and 27B open, Plus and Max hosted | Open weights, Apache 2.0 |
| Qwen3.7 | May–Jun 2026 | API-only generation; no open weights were ever released | API-only |
| Qwen3.8 | Aug 2026 | First open release of a Max-class flagship; the current open generation | Open weights (with licence exceptions) |
| Qwen4 | 16 Nov 2026 | Qwen4-Max, the new flagship; 1M-token context; API-only | API-only |
Read the table with two warnings. First, the version numbers are not a quality ranking. Qwen3.5, Qwen3.6, Qwen3.7 and Qwen3.8 are sub-generations of the Qwen3 era, and the jump from one to the next is smaller than the jump from Qwen2.5 to Qwen3. A bigger number does not automatically mean a better model for your job — Qwen3.8-27B is the practical default for local use not because 3.8 is the biggest number, but because it is the newest open generation in a size that fits on real hardware. Second, the open/API-only column is the column that matters most. In most generations, the Max, Plus and Flash tiers are hosted-only: you can call them through Alibaba Cloud’s Model Studio, but you cannot download them. The exception that proves the rule is Qwen3.8, where the Max-class model itself was opened — the first time Alibaba has ever released a Max-class flagship as open weights. That is why Qwen3.8, not Qwen4, is the generation you can actually download today.
The timeline also explains why older Qwen models still appear in so many tutorials. Qwen2.5 and Qwen3 were widely deployed, and their smaller sizes run on modest hardware, so they remain in use long after their successors arrive. Nothing in this guide treats an older generation as obsolete on the day its successor ships. The question is always what the model is for, and the verdict section at the end answers it by use case rather than by date.
Qwen4-Max: The New Flagship, Released 16 November 2026

Qwen4-Max is the first full Qwen4 model, and it arrived on 16 November 2026 — one day before this guide was checked. Its model ID is qwen4-max. It is a preview-to-release story with a short fuse: a preview version, qwen4-max-preview, ran from 10:00 on 11 November to 24:00 on 25 November 2026 (UTC+8), available through Alibaba’s Token Plan only, and it is retired after that window. Do not build on the preview ID; it is a temporary artefact, not a product. The full release on 16 November is the model that matters. Our news post on the launch gives the day-one story; this section is the durable reference. Alibaba’s Qwen4-Max launch, covered on 16 November 2026.
What is known about the model itself is as follows, as reported from Alibaba’s own Model Studio pages by multiple secondary reads on 16 November 2026 — the direct Model Studio page was not opened first-hand during this research, so the figures below are verified-by-secondary-read, not first-hand, and no figures beyond them are added here. Qwen4-Max has a 1,000,000-token context window. It takes text, image and video as input and produces text as output. Reasoning is on by default. It is API-only: there are no downloadable weights, and Alibaba’s documentation marks fine-tuning and custom deployment as unsupported. It is available in six listed regions: Beijing, Singapore, Frankfurt, Virginia, Tokyo and Hong Kong.
The price, for the Singapore and international region, is $1.50 per million input tokens and $4.50 per million output tokens, checked 17 November 2026. Beijing, Hong Kong, Frankfurt, Virginia and Tokyo list slightly lower prices — $1.413 per million input tokens and $4.24 per million output tokens. Cached input is reported at $0.075 per million tokens. That pricing is about 25% below Qwen3.8-Max, which costs $2.00 per million input tokens and $6.00 per million output tokens in Singapore. New Singapore accounts are reported to receive 1 million free tokens for 90 days. No Token Plan subscription price is stated in this guide, because the reported figure for a personal plan was not verified on Alibaba’s pricing page.
One figure this guide deliberately does not print is a parameter count for Qwen4-Max. Secondary reports disagree about whether Alibaba has published one at all: Alibaba’s launch announcement as quoted by some outlets describes a very large mixture-of-experts model, while another report states plainly that Alibaba has not published a parameter count. Until the figure is confirmed first-hand, printing it as settled fact would be guessing, and this guide does not guess. The same discipline applies to benchmarks. Secondary reports cite an Artificial Analysis Intelligence Index score of 43 for Qwen4-Max, against 34 for Qwen3.8-Max — but that score was not verified on Artificial Analysis itself, and Alibaba’s own benchmark claims are vendor-reported. No benchmark score in this guide is presented as an independent verdict, and no head-to-head verdict against Claude, GPT or Gemini is stated, because no independent leaderboard result was verified first-hand.
The wider context is worth knowing. At the Apsara Conference on 22 September 2026, Alibaba said Qwen4 was “currently in training”, with a roadmap running on to Qwen4.5 and Qwen5. A Qwen4 teaser poster went out around 6 November 2026, which we covered at the time. Smaller open-weight Qwen4 models have been promised as “coming to open source” — but with no names, no sizes, no dates and no licence given. This guide names none of them, because none exists to name: as of 16–17 November 2026 there is no Qwen4 repository on Hugging Face, and no model called “Qwen 4 27B” exists. Any third-party listing of one should be treated with suspicion. The honest summary of Qwen4 today is short: one model, one day old, API-only, a million tokens of context, and a promise of smaller open models to come.
Qwen3.8: The Generation You Can Actually Download

If Qwen4-Max is the flagship you can only rent, Qwen3.8 is the generation you can actually own. Released in August 2026, it is the current open generation — and it contains the most significant open-weight release in Qwen’s history. Three models define it, and they serve three different readers.
The headline is Qwen3.8-2.4T-A95B, also called Qwen3.8-Max, whose open weights were released on 12 August 2026. It is a sparse mixture-of-experts model with 2.4 trillion total parameters and about 95 billion active per token, and a 1-million-token context in its hosted form. It is the first time Alibaba has ever released a Max-class flagship as open weights. There is a catch, and it is a licence catch, not a technical one: Qwen3.8-2.4T-A95B is released under a custom Qwen3.8-Max licence, not Apache 2.0, and Hugging Face tags it license: other. Before any commercial use, check the licence file on the model’s Hugging Face card — the exact restriction wording is reported differently by different sources, and this guide does not quote a revenue threshold or any other specific restriction, because the sources disagree on the wording. The model card is here: Qwen3.8-2.4T-A95B on Hugging Face.
The practical pick is Qwen3.8-27B, released as open weights on 14 August 2026. It is a dense model of 27.78 billion parameters with native vision — text, image and video in, text out — and a native context of 262,144 tokens. That context figure needs a careful reading: the 1-million-token figure sometimes quoted for this model belongs to the hosted or YaRN-extended version, not to the downloaded weights, whose native context is 262,144 tokens (extendable toward 1 million via YaRN). It is released under Apache 2.0, which makes it the default local Qwen for most people: a current-generation model, a permissive licence, and a size that fits on a single 24 GB graphics card at roughly 17–18 GB in 4-bit quantisation — an approximate, rule-of-thumb figure, not a measured benchmark. If you want one Qwen to download today, this is the one.
The third model is the most forward-looking. Qwen3.8-Flash-Next, released on 26 August 2026, is a mixture-of-experts model with 125 billion total parameters and only 6 billion active per token, a 51-billion-parameter n-gram embedding table, and a 262,144-token native context extensible to 1 million. It is described as a preview of the Qwen4 architecture — which makes it the closest thing to a bridge between the open Qwen3.8 generation and the closed Qwen4 flagship. Its licence is qwen-community-1.0, not Apache 2.0, so the same advice applies: read the licence file before commercial use.
The rest of the Qwen3.8 family is hosted. Qwen3.8-Flash is the fast hosted tier. Qwen3.8-Omni-Flash arrived on 18 September 2026, alongside Qwen3.8-Omni-Flash-Realtime and Qwen3.8-LiveTranslate, a real-time interpretation model whose average lagging fell from 2.8 seconds to 2.3 seconds, according to the Qwen blog. The hosted Qwen3.8-Max API costs $2.00 per million input tokens and $6.00 per million output tokens in Singapore, checked 17 November 2026 — the price Qwen4-Max undercuts by about 25%. The pattern of the whole generation is clear: the open weights are the story, the hosted tiers are the business, and the licence is the detail that decides what you can actually do with each model.
Qwen3.5 and Qwen3.6: The Generations Still Worth Knowing

Qwen3.5 and Qwen3.6 are the generations most readers skip — and that is a mistake, because they are still the right answer for a large number of real deployments. Both are fully open under Apache 2.0, both are natively multimodal, and both cover size ranges that the newer generations do not replicate exactly.
Qwen3.5 arrived in February and March 2026. The first open checkpoint, Qwen3.5-397B-A17B, was released on 16 February 2026; the 122B-A10B, 35B-A3B and 27B sizes followed on 24 February; and the small 0.8B, 2B, 4B and 9B models arrived on 2 March 2026. It is natively multimodal — text, image and video in — and it uses a Gated DeltaNet hybrid attention design, with a 256,000-token context. Qwen3.5-Omni, the all-modalities variant, followed on 31 March 2026. The range matters: at the top, a 397-billion-parameter model for serious deployments; at the bottom, sub-10-billion models that run on laptops and edge devices. That small end is the reason Qwen3.5 still matters — Qwen3.8’s smallest open model is far larger, and a laptop cannot run a 27-billion-parameter model comfortably.
Qwen3.6 followed in April 2026: Qwen3.6-35B-A3B on 16 April and Qwen3.6-27B on 22 April, with hosted Qwen3.6-Plus from 2 April and a Qwen3.6-Max-Preview from 18 April. The open weights are Apache 2.0, with a 262,144-token context. It is best understood as a refinement generation — the Qwen3 architecture, polished, in the two sizes that had proved most useful — rather than a reinvention. For teams with an existing Qwen3.6 deployment, there is no urgent reason to move: the model is current, the licence is permissive, and the hosted tiers are still served.
The honest way to think about these two generations is as infrastructure, not as news. They are the models already running in products, on servers, and on hardware that cannot be upgraded overnight. Newer is not automatically better for a deployment that works: a Qwen3.5-4B model running on a laptop, or a Qwen3.6-27B model serving an internal tool, does not become a bad model because Qwen3.8 exists. It becomes a model with a known replacement path — and the replacement path is what the verdict section at the end of this guide maps out.
Qwen3: The Release That Changed Everything

Qwen3, announced on 28 April 2025 with repositories published the next day, is the release that changed what Qwen is. Everything before it was a conventional model family; everything after it carries Qwen3’s central idea. That idea is simple to state and hard to build: one model that can both think and answer, with a switch between the two.
The release covered eight open-weight models: dense models at 0.6B, 1.7B, 4B, 8B, 14B and 32B parameters, and mixture-of-experts models at 30B-A3B and 235B-A22B. All of them are Apache 2.0. They were pre-trained on about 36 trillion tokens and cover 119 languages and dialects. The native context was 32,000 tokens, 128,000 in the initial release, and the 2507 updates of July and August 2025 raised it to 256,000 tokens, extendable toward 1 million. The 2507 updates also split the line into Instruct and Thinking variants — a sign of how central the thinking question had become.
The thinking mode is what absorbed QwQ. In Qwen3, a single model runs in thinking mode — working through a problem step by step before answering — or in non-thinking mode, answering directly, and the thinking budget is adjustable. That unification is documented in the Qwen3 Technical Report (arXiv:2505.09388), which this guide read in full. The report’s own comparison is a vendor claim, and it is stated as one: Qwen3-32B in thinking mode outperforms QwQ-32B on 17 of 23 benchmarks. Whether or not that figure impresses you, the direction is clear — the standalone reasoning line was no longer needed once reasoning became a mode inside the general model.
Qwen3 also produced the architecture that shaped the generations after it. Qwen3-Next-80B-A3B, released in September 2025, is an ultra-sparse mixture-of-experts model with 80 billion total parameters and only 3 billion active per token, using a Gated DeltaNet and gated-attention hybrid design, with a context of 262,000 to 1 million tokens. Its architecture later informed Qwen3.5 and beyond — and its descendant, Qwen3-Coder-Next, is the current head of the coding line. Qwen3 is the generation where Qwen stopped being a collection of models and became a system: one architecture, one licence, one thinking mechanism, in sizes from a phone to a data centre.
QwQ Explained: Alibaba’s Reasoning Experiment — and Why It Ended

QwQ is the model that put Qwen on the front page in the winter of 2024–25, and it is also the model most likely to be misunderstood today — because the right way to describe it is in the past tense. QwQ was an experiment in reasoning, it worked, and then it was absorbed into the main line. There is no QwQ successor, and there has been no new QwQ release since March 2025.
The story runs in two releases. QwQ-32B-Preview arrived on 28 November 2024 as an experimental open-weight model. QwQ-32B followed on 5–6 March 2025: a dense 32.5-billion-parameter model built on the Qwen2.5-32B base, trained with supervised fine-tuning and reinforcement learning, which produced its answers inside explicit <think> blocks — a visible chain of thought the reader could watch unfold. Its context was 131,072 tokens, extended from a 32,000-token base via YaRN, and it was released under Apache 2.0. Alibaba positioned it against DeepSeek-R1 and OpenAI’s o1-mini — that positioning is the vendor’s, not an independent verdict, and this guide repeats it as positioning only.
The visual-reasoning branch ran alongside it. QVQ-72B-Preview, a visual-reasoning experiment, arrived in December 2024, and QVQ-Max followed on 28 March 2025 as a hosted-only model. Both were experiments in the same direction: what happens when a model is trained to reason out loud, in text or in images, before it answers.
So why did it end? Because the experiment succeeded too well to stay separate. Once Qwen3 put thinking mode inside every general model — switchable, with an adjustable thinking budget — a standalone reasoning line had nothing left to offer that the main line did not. The Qwen3 Technical Report says so in its opening section: reasoning was consolidated into thinking mode. QwQ-32B remains downloadable and usable, and its Apache 2.0 licence means it will keep running in old deployments and tutorials for years. But nobody should choose it today as a reasoning model. The successor to QwQ is not a model called QwQ-2. It is the thinking switch inside Qwen3, Qwen3.5, Qwen3.6 and Qwen3.8 — and, by default, inside Qwen4-Max, where reasoning is on unless you turn it off. Search interest in “qwq” is still high, and this section exists because of that interest: the answer the searches need is that the line is finished, the idea survived, and the idea now lives inside every current Qwen.
Qwen-Coder: The Coding Line, From Qwen2.5-Coder to Qwen3-Coder-Next

The coding line is Qwen’s most practical branch, and it has a clear head today: Qwen3-Coder-Next, released in February 2026. Getting there took three generations, and each one is still worth knowing, because the older coders are the ones running in most local setups.
Qwen2.5-Coder, released between September and November 2024, came in six sizes — 0.5B, 1.5B, 3B, 7B, 14B and 32B — all dense, with base and instruct versions at every size. It was pre-trained on 5.5 trillion code tokens, a vendor figure, with a 32,000-token default context extendable to 131,072 via YaRN, and fill-in-the-middle support for completing code in the middle of a file. It is Apache 2.0, with one exception: Qwen2.5-Coder-3B is listed on its Hugging Face card under a qwen-research licence, not Apache 2.0. The 32B instruct model was reported at 92.7% on HumanEval Pass@1 — a vendor and self-reported figure, stated here as reported, not as an independent result. For readers choosing between coding assistants more broadly, our comparison of AI coding assistants sets out how the Qwen coders sit against the rest of the field.
Qwen3-Coder arrived on 22 July 2025 in two sizes: 30B-A3B and 480B-A35B, both mixture-of-experts, both instruct-only, both Apache 2.0. It was built for agentic coding — the kind of coding where the model plans, writes, runs and fixes code in a loop, rather than completing a single function. It launched alongside Qwen Code, Alibaba’s terminal coding agent. One detail matters for anyone migrating from QwQ-style reasoning: Qwen3-Coder models do not emit <think> blocks. They are non-thinking only, by design, because an agent that pauses to think out loud on every step is an agent that wastes its context.
Qwen3-Coder-Next, released on 2–3 February 2026, is the current head of the line. It is an 80B-A3B mixture-of-experts model built on the Qwen3-Next-80B-A3B base, designed for coding agents and local development, with a 256,000-token context and an Apache 2.0 licence. No Qwen3.8 Coder variant had shipped as of the sources checked for this guide, and no “Qwen4 Coder” exists. Which coder for which machine? The 0.5B–7B Qwen2.5-Coder sizes run on a laptop; the 14B and 32B sizes suit a workstation; Qwen3-Coder-Next at 80B total and 3B active is the efficient choice for an agent on a single good GPU; and the 480B Qwen3-Coder is a data-centre model, not a local one.
The tool around the models matters as much as the models. Qwen Code is an open-source terminal coding agent, originally based on Google’s Gemini CLI, and its 1.0 Preview — released on 11 November 2026 — folds in the agent capabilities of Qoder, Alibaba’s other coding agent. It ships as native command-line packages for macOS, Windows and Linux, with no separate Node.js install required. Our news post on the 1.0 Preview gives the day-one detail; the point for this guide is that the coding line is no longer just weights on Hugging Face. It is weights plus an agent, and the agent is where the line is heading. Alibaba folds Qoder into open-source Qwen Code 1.0, covered 11 November 2026.
Seeing, Hearing, Speaking: Qwen3-VL, Omni, and the Multimodal Branches

Qwen’s general models can already see — Qwen3.8-27B takes images and video as input — but the dedicated vision and omni lines go further. They are separate branches with their own releases, and they are where Qwen’s breadth is most visible.
The vision line runs Qwen-VL, then Qwen2.5-VL in January and February 2025 (3B, 7B, 32B and 72B), then Qwen3-VL in September and October 2025 (2B, 4B, 8B and 32B, plus 30B-A3B and 235B-A22B mixture-of-experts variants). These are dedicated vision-language models: the right choice when the job is reading documents, understanding screens and interfaces, or answering questions about images, rather than chatting about them. Published weights are Apache 2.0, but the licence should be verified on each individual model card, because the family has exceptions elsewhere.
The omni line is the more ambitious idea: one model that takes text, image, audio and video in, and produces text or speech out. Qwen2.5-Omni arrived on 27 March 2025 as a 7-billion-parameter model built on a Thinker–Talker design — one part that understands, one part that speaks — under Apache 2.0. Qwen3-Omni followed in September 2025, and Qwen3.8-Omni-Flash arrived on 18 September 2026, alongside a realtime variant. The visual-reasoning cousin, QVQ, is covered in the QwQ section above: QVQ-72B-Preview in December 2024 and the hosted-only QVQ-Max in March 2025.
Two more branches complete the picture. Qwen-Audio and Qwen3-TTS cover speech: text-to-speech including voice design and cloning, with demonstrations on Hugging Face Spaces. And Qwen3.8-LiveTranslate, released on 18 September 2026, is real-time interpretation — speech in one language, speech out in another, with average lagging down from 2.8 seconds to 2.3 seconds, according to the Qwen blog. The practical point is that “multimodal” in Qwen is not one feature. It is a set of branches, each with its own release history, and the right one depends on whether you need to read an image, hear a voice, or translate a conversation as it happens.
Qwen-Image and the Specialist Models Nobody Maps

Most Qwen maps stop at chat, code and vision. The family is wider than that, and the specialist branches are where the most useful surprises live — for image generation, for search, for translation, and for safety.
Qwen-Image is the image line. The foundation model arrived on 4 August 2025: a 20-billion-parameter MMDiT image model noted for strong text rendering inside generated images. Qwen-Image-Edit followed for image editing. Qwen-Image-2.1, released on 20 September 2026 as open source, unifies text-to-image generation and editing in a 7-billion-parameter visual-generation component, with native support for transparent images. Qwen-Image-3.0, released on 22 July 2026, is hosted only, in Pro and Turbo API variants, with prompts of up to 4,500 tokens in 12 languages. Qwen-Image-Edit-2511, at around 20 billion parameters, is reported as Apache 2.0. The pattern is the same as the general line: the newest open image model sits one step behind the newest hosted one.
The search branch is Qwen3-Embedding and Qwen3-Reranker, the pair that powers retrieval-augmented generation: the embedding model turns text into searchable vectors, the reranker orders the results. They come in 0.6B, 4B and 8B sizes, with the 0.6B model reaching 1,024 dimensions and 32,000-token inputs, across more than 100 languages, under Apache 2.0. For anyone building a search system on Qwen, these are the models that matter most — and the most commonly overlooked.
Two more specialists deserve a line each. Qwen-MT, also called qwen-mt-turbo, is the translation model: 92 languages, covering more than 95% of the world’s population, according to the Qwen blog — a vendor claim, stated as one. Qwen3Guard is the family’s first safety guardrail model, classifying prompts and responses with risk levels, described on the Qwen blog. Together with Qwen-Audio and Qwen3-TTS from the previous section, they complete the picture: Qwen is not a chat model with extras. It is a toolkit, and the chat model is one tool in it.
Open or Closed? Licences and Access, Decoded

“Is Qwen open source?” is the most-asked question about the family, and the honest answer has three parts. First, most Qwen open weights are released under Apache 2.0, one of the most permissive licences in software. Second, there are exceptions, and they sit exactly where the money is. Third, open weights are not the same as open source — and the difference matters for anyone building a business on Qwen.
The default is Apache 2.0. Qwen3, Qwen3.5, Qwen3.6, Qwen3.8-27B, Qwen3-Coder, Qwen3-Coder-Next, Qwen2.5-Coder (except one size), Qwen2.5-Omni and Qwen3-Embedding are all Apache 2.0. That licence lets you use the weights commercially, modify them, and distribute them, with only attribution and a copy of the licence required. It is the reason Qwen is the default open-weight choice for so many developers, and the reason the family appears in so many products.
The exceptions are three. Qwen3.8-2.4T-A95B — the open Max flagship — is under a custom Qwen3.8-Max licence, not Apache 2.0, and Hugging Face tags it license: other. Qwen3.8-Flash-Next is under qwen-community-1.0, not Apache 2.0. And Qwen2.5-Coder-3B is listed on its Hugging Face card under qwen-research, not Apache 2.0. For the two custom licences, this guide states only that they are not Apache 2.0 and that the licence file on Hugging Face must be checked before commercial use. It does not quote any revenue threshold or restriction wording, because the sources checked disagree on the exact wording. That disagreement is itself the lesson: when a model’s licence is custom, the licence file is the authority, not a summary in a guide — including this one.
Then there is the hosted side. Max, Plus and Flash tiers are hosted-only in most generations: you call them through Alibaba Cloud’s Model Studio, you pay per token, and you never see the weights. Qwen3.7 was API-only in its entirety. Qwen4-Max is API-only. The practical rule is simple: if a tier name ends in Max, Plus or Flash, assume it is hosted-only unless the release notes say otherwise — and Qwen3.8-Max is the one release that says otherwise.
Finally, the phrase “open-weight” itself. Open weights mean you can download the trained parameters and run the model. They do not mean the training data is published, the training code is published, or the model is open source in the sense the Open Source Initiative defines it. Qwen is open-weight, not open source, and the distinction matters commercially: you can build on the weights, but you cannot audit how they were made. For the wider context on open-weight models and the retrieval systems that use them, our explainer of BERT, Llama, Whisper and RAG covers the Llama side of the open-weight world and the embedding and retrieval ideas behind Qwen3-Embedding. The checklist for any Qwen model is short: read the licence file on its Hugging Face card, confirm whether the tier you want is open or hosted, and never assume that what is true of Qwen3.8-27B is true of Qwen3.8-2.4T-A95B.
How to Try Every Qwen Model: Chat, API, or Your Own Machine

There are three ways into Qwen — chat, API, or your own machine — and the right one depends on what you want to do, not on which is most impressive. Start with the easiest, and move down the list only when the easier route stops being enough.
The chat route is Qwen Chat at chat.qwen.ai, free to try, with the Qwen app on phone and desktop alongside it. It runs the latest models behind a simple chat box, with image generation, Deep Research, Web Dev, video generation, search and multimodal understanding, according to the Qwen homepage. It is the right first step for anyone who wants to know what Qwen feels like: ask it questions, upload an image, try the thinking mode, and see whether the answers are useful. This guide states no free-tier limits and no account requirements, because neither was verified on chat.qwen.ai during this research.
The API route is Alibaba Cloud Model Studio, also known as DashScope. It serves every hosted tier — all the Max, Plus and Flash models, and Qwen4-Max — through an OpenAI-compatible API format, which means code written for OpenAI’s API can often be pointed at Model Studio with a changed endpoint. Access is through a Token Plan subscription; no subscription price is stated here, because the reported figure was not verified on Alibaba’s pricing page. Per-token prices for the flagship models are given in the Qwen4-Max and Qwen3.8 sections above, date-stamped checked 17 November 2026. Third-party hosts — OpenRouter, Together, Fireworks and the Vercel AI Gateway — also serve Qwen models; Qwen4-Max is listed as alibaba/qwen4-max on Vercel, as reported on 15 November 2026. For the wider question of how API access compares across providers, our explainer of the OpenAI API covers the patterns — keys, pricing, rate limits — that apply to Model Studio as much as to any other provider.
The local route is the open weights, from Hugging Face (the Qwen organisation: 469 models, 37 collections) and ModelScope, run through Ollama, vLLM, llama.cpp or LM Studio. The command for the default pick is ollama run qwen3.8. Size picks by hardware, all approximate: Qwen3.8-27B at roughly 17–18 GB in 4-bit quantisation, fitting a single 24 GB graphics card; Qwen3-30B-A3B at roughly 19 GB in 4-bit quantisation, a fast decoder for its size; and the small Qwen3.5 models at 0.8B to 9B for laptops and edge devices. Those figures are rules of thumb, not measured benchmarks, and actual memory use depends on quantisation and context length. Start small, confirm the model runs, and only then move up in size — a 27-billion-parameter model that runs is worth more than a 235-billion-parameter model that does not fit.
For coders, there is a fourth route: Qwen Code, the open-source terminal coding agent, installable as a native command-line package on macOS, Windows and Linux with no separate Node.js install. It pairs naturally with the Qwen3-Coder models, and its 1.0 Preview folds in Qoder’s agent capabilities. The order of operations for a new user is: try Qwen Chat first, move to the API when you need to build with it, run weights locally when you need control or privacy, and install Qwen Code when the job is writing software rather than chatting about it.
Qwen vs the Rest: DeepSeek, Llama, Kimi, and the Closed Frontier

Qwen does not exist in a vacuum, and the comparison that matters most is with the other open-weight families — because that is where Qwen actually competes. The closed frontier models from OpenAI, Anthropic and Google are a different question, and this section keeps the two comparisons apart.
Against DeepSeek, Qwen’s advantage is breadth. DeepSeek has one famous reasoning model line and a small number of releases; Qwen has thirteen branches, sizes from under a billion to trillions of parameters, and a new release every few months. DeepSeek-R1 was the model QwQ was positioned against in March 2025, and the two were genuine rivals for a season — but DeepSeek has not matched Qwen’s release cadence since, and Qwen’s reasoning now lives inside its general models rather than in a separate line. Against Llama, from Meta, Qwen’s advantage is the licence and the range: Qwen’s open weights are mostly Apache 2.0, while Llama carries Meta’s custom community licence, and Qwen ships far more sizes and specialist models. The Llama side of the open-weight world is covered in our explainer of the open-weight and retrieval landscape, linked in the licences section above. Against Kimi, from Moonshot AI, Qwen’s advantage is again breadth and availability: Kimi’s models are strong, but they are fewer, and Qwen’s distribution through Hugging Face, ModelScope, Ollama and every major third-party host is wider.
Where Qwen does not win is equally clear. Its flagship is closed. Qwen4-Max is API-only, with no downloadable weights and no fine-tuning, while the open line sits one generation behind at Qwen3.8. For a developer who wants the best Qwen and wants to run it, that is a real gap — the best model and the downloadable model are not the same model. The second gap is evidence: no independent benchmark result for Qwen4-Max was verified first-hand during this research, and the scores that circulate are secondary or vendor-reported. Any guide that declares a winner between Qwen4-Max and Claude, GPT or Gemini on the basis of those scores is declaring a verdict it has not earned. The third gap is documentation: with 469 models in the Qwen organisation, finding the right one is genuinely hard, and the family’s own naming — Max, Plus, Flash, Next, Omni, VL, MT, Guard — is a vocabulary that takes a guide like this one to decode.
The closed frontier is a different comparison, and it should be stated honestly. Claude, GPT and Gemini are closed models with large research organisations behind them, and their makers publish more evaluation detail than Alibaba does for Qwen. For background on each, the sibling explainers in this series are the right next reading: our explainer of the OpenAI model family, our explainer of the Claude model family, our explainer of the Gemini model family, and our explainer of Grok. For the decision itself — which model to choose for a given job, and what speed, cost, reasoning depth and context actually trade off against each other — our guide to choosing an AI model sets out the framework. The short version is that Qwen wins on breadth, sizes and licence among open weights, and loses to the closed frontier on one point only: the best Qwen is not the Qwen you can download.
Which Qwen Should You Pick? A Verdict by Use Case — and What to Watch Next

There is no single best Qwen. There is a best Qwen for each job, and the verdict below is by use case — the only honest way to end a guide about a family of 469 models.
| Use case | Pick | Why |
|---|---|---|
| Casual chat | Qwen Chat at chat.qwen.ai | Free to try, latest models, no setup |
| Running a model locally | Qwen3.8-27B | Current open generation, Apache 2.0, fits a single 24 GB GPU |
| Coding, local or agentic | Qwen3-Coder-Next + Qwen Code | Current coder line with the 1.0 Preview agent |
| Vision and documents | Qwen3-VL | Dedicated vision line, sizes from 2B to 235B |
| Images | Qwen-Image-2.1 (open) or Qwen-Image-3.0 (hosted) | Open editing and generation, or the hosted Pro/Turbo tiers |
| Search and RAG | Qwen3-Embedding + Qwen3-Reranker | Purpose-built retrieval pair, Apache 2.0 |
| API production | Qwen4-Max, or Qwen3.8-Max for the open-weight path | Newest flagship, or the Max you can download |
| Laptop or edge | Qwen3.5 small sizes (0.8B–9B) | Small, open, Apache 2.0 |
The limits belong in the same breath as the recommendations. Qwen4-Max is one day old as this guide is checked, and its benchmark evidence is secondary and vendor-reported; treat early claims about it with the scepticism any one-day-old model deserves. The open line lags the flagship by a generation, and the smaller open Qwen4 models Alibaba has promised have no names, dates or licences yet — they are a promise, not a product. The licences have exceptions at exactly the top of the range, so the licence file must be checked per model. And the family’s size is itself a hazard: with 469 models, the wrong Qwen is always one click away, and the version numbers will not protect you.
What to watch next is short, and every item on it is unconfirmed. The smaller open-weight Qwen4 models, promised but unnamed — the moment one appears on Hugging Face, the open line catches up a generation. The Qwen4.5 and Qwen5 roadmap mentioned at the Apsara Conference in September 2026, with no figures attached in this guide because the reported figures were not verified. And the image line’s next step, with Qwen-Image-3.0 already hosted and the open line at Qwen-Image-2.1. None of these has a date. The family moves every few months, and the check that matters is the one this guide performed on 17 November 2026: read the release notes, read the licence file, and do not trust a guide — including this one — past its checked date. Checked 17 November 2026.
Sources
All sources below were read on 17 November 2026, except where noted. Facts, dates, prices and licence details in this guide come from these pages. No competitor or comparison-site pages were used as sources. Prices checked 17 November 2026.
- Qwen — homepage (features; Latest Research: Qwen-Image-2.1, 20 Sep 2026; Qwen3.8-LiveTranslate; Qwen3.8-Omni-Flash) — https://qwen.ai/
- Qwen official blog index (Qwen3Guard, Qwen-Image-Edit, Qwen-Image, Qwen-MT entries) — https://qwenlm.github.io/blog/
- Hugging Face — Qwen organisation card (469 models, 37 collections; links chat.qwen.ai) — https://huggingface.co/Qwen
- Qwen3 Technical Report (arXiv:2505.09388; thinking/non-thinking unification; Qwen3-32B vs QwQ-32B, 17/23 benchmarks, vendor claim) — https://ar5iv.labs.arxiv.org/html/2505.09388
- Alibaba Cloud Model Studio — Qwen4-Max model page, release notes and Token Plan page (read via secondary reads, cross-checked across multiple sources: The Latent Daily, The Model Atlas, Apidog, enjoy.dev, CellCog — consistent on 16 Nov 2026 date, model ID qwen4-max, $1.50/$4.50 pricing, 1M context, preview window 11–25 Nov 2026; direct page not opened first-hand)
- QwenLM GitHub repositories — Qwen3.8, Qwen3.5 and Qwen3.6 release records (read via secondary reads, cross-checked: thinkfacility.com, try-qwen-ai.com, blog.lqd3-solutions.ai, codersera.com, aixjo.com — consistent on dates and licences)
- Hugging Face Qwen model cards — Qwen3.8-27B, Qwen3.8-2.4T-A95B, Qwen3-Coder-Next, Qwen2.5-Coder (read via secondary reads: urhoba.net, detartech.com, refusalrate.com, xinference.co — consistent; licence conflicts flagged in the guide; direct open of individual cards was not possible in this session)
- Qwen3.8-2.4T-A95B model card — https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
- Qwen Code documentation (qwenlm.github.io/qwen-code-docs) — Qwen Code Weekly, 29 Oct 2026
- Qwen official blog — “Embracing the Power of Reinforcement Learning” (QwQ-32B) — via https://qwenlm.github.io/blog/
- GER news — “Alibaba’s Qwen4-Max Arrives as a 24/7 AI Teammate” (16 Nov 2026) — https://geroundtable.com/qwen4-max-launch/
- GER news — “Alibaba Folds Qoder Into Open-Source Qwen Code 1.0” (11 Nov 2026) — https://geroundtable.com/qwen-code-1-0-qoder-open-source/
- GER news — “Alibaba Teases Qwen4 With a Riddle and No Release Date” — https://geroundtable.com/qwen4-teaser-becoming-begins/


0 Comments