Qwen by Alibaba in 2026: Open Weights Are Back at the Top

AI models in this article
While managers pay for ChatGPT Plus and Claude Pro, Alibaba has built one of the largest AI ecosystems in the world – a free chat app, open models and API prices that undercut the competition several times over. Qwen (pronounced “chwen”, from 通义千问 – “A Thousand Questions”) had overtaken every Western rival in downloads by mid-2026 and became a tool every manager should know. In the spring it looked as if the ecosystem was drifting toward the OpenAI and Anthropic playbook: the 3.7 flagships shipped closed, and key developers walked out. In August, Alibaba released three models and turned back. The flagship Qwen3.8-Max comes with a promise of open weights, and Qwen3.8-27B, under Apache 2.0, runs on a single consumer GPU.
The pitch for managers is simple: Qwen is a full-featured AI assistant you can use for free at Qwen Chat – no subscription, no cap on requests. By April 2026, Qwen models were closing in on a billion downloads, more than half of all open-model downloads worldwide. The consumer app has over 200 million monthly active users, third in the world behind ChatGPT and DouBao.

What makes Qwen different?
What sets Qwen apart from most competitors is scale and flexibility. Alibaba publishes a large share of its models openly on Hugging Face, so any company can download them and run them on its own servers. Kazakhstan’s national language model, Oylan, is one of the projects built on top of Qwen. This summer, open weights stopped being a mid-tier feature: Alibaba announced it would also release the weights of its flagship, Qwen3.8-Max. The 3.7 series, meanwhile, stays closed and API-only.
- Most Qwen models are available to download and run on your own hardware. For companies with strict data-security requirements, the open models – Qwen3.8-27B, Qwen3.8-Flash-Next, Qwen 4 Coder and the entire Qwen 3.5 family – remain among the best in the world.
- The flagships use a mixture-of-experts design: each request is handled by only the relevant group of experts, a small active slice of the model. In practice, a model that runs on a laptop now delivers answers that used to require expensive server hardware.
- Qwen3 text models support 119 languages and dialects; the newest versions support 201.
- Text, images, video, audio and speech generation all live in one ecosystem.
What Qwen can do in September 2026
Today’s Qwen lineup covers almost every kind of task a modern company runs into:
- A flagship with promised open weights. Qwen3.8-Max (GA on August 3, 2026) is a sparse MoE with 2.4 trillion parameters, about 95 billion of them active on each pass. 1M-token context; takes text, images and video as input. $2.00/$6.00 per 1M tokens.
- Local deployment with vision. Qwen3.8-27B (August 13–14) is a dense 27–28B model with a vision encoder, 262K context, an Apache 2.0 license and roughly 17 GB at Q4_K_M quantization: one GPU, or a laptop with 16 GB of memory.
- Twelve times cheaper. Qwen3.8-Flash-Next (August 26) is an open preview of the Qwen4 architecture: 125B parameters, 6B active per token, about $0.16/$0.47 per 1M tokens.
- Next-generation agents. Qwen 3.7 Max (May 2026) is built as an “agentic model” – it can autonomously run chains of 1,000+ tool calls in a single session without losing quality. 1M-token context and the lowest hallucination rate among flagship models (22.9%).
- A multimodal agent. Qwen 3.7 Plus (GA on June 1, 2026) understands text, images and video, sees the screen and returns the coordinates of interface elements (GUI agent grounding). 1M-token context.
- A new bar for coding. Qwen 4 Coder 32B-A3B (June 2, 2026) is the first open model to hit 82% on SWE-Verified. MoE architecture: 32B parameters, only 3B active – it runs on a Mac with 24 GB of memory at 58 tokens/sec. Apache 2.0 license.
- Long documents. Qwen3.6-Plus holds up to 1 million tokens – about 2,000 letter-size pages in a single request. Load a full annual report, a customer-ticket archive or an entire code repository and ask questions across all of it.
- Hard reasoning. Qwen3.7-Max in chain-of-thought mode scores 92.4% on GPQA Diamond – ahead of Claude Opus 4.6 Max (91.3%) and K2.6 Thinking (90.5%). On math (Apex Math), it scores 44.5% against Claude Opus 4.6 Max’s 34.5%.
- Agentic coding. Qwen3.6-35B-A3B works alongside Claude Code, Qwen Code and similar tools, solving tasks at Claude Sonnet 4.5’s level for a fraction of the price. Open license – you can host it in-house.
- Voice and video in one model. Qwen3.5-Omni understands text, images, audio and video within a single architecture. It powers “Audio-Visual Vibe Coding”: you describe the task out loud and show it on screen, and the model writes the code.
- Image and speech generation. Qwen-Image-2512 produces photorealistic images with accurate multi-line text – rare for generative models. Qwen3-TTS synthesizes speech with voice cloning for narration and voice assistants.
- Third on the LMArena vision leaderboard. Qwen3.8-Max scores 1,300±8, behind only two Anthropic models. The confidence intervals of the top nine overlap, so this is a tight pack, and the top spots are separated by less than the margin of error.
Qwen’s main models (September 2026)
By September 2026, Alibaba had assembled one of the broadest model lineups in the industry.

Qwen3.8-Max – the flagship Alibaba promised to open (August 2026)
Qwen3.8-Max went GA on August 3, 2026, and received a post-training update tagged 0902 on September 2–3. It isn’t in our benchmark yet – the figures below come from the vendor and independent leaderboards.
The architecture is a sparse mixture of experts: 2.4 trillion parameters in total, about 95 billion active on each pass. 1M-token context, with responses of up to 131,072 tokens. It takes text, images and video and returns text; reasoning is on by default. Tool calling and structured output via a JSON schema in response_format are supported. Pricing per 1M tokens: $2.00 input, $6.00 output, $0.25 for cache reads.
What Alibaba itself claims:
- OSWorld-Verified – 86.1, versus 83.2 for GPT-5.6 Sol Max, 85.0 for Claude Fable 5 and 76.2 for Gemini 3.1 Pro
- PaperBench – 93.0, the highest published result on that benchmark
These are the vendor’s numbers from the vendor’s own runs; treat them accordingly.
Independent leaderboards confirm the headline. On LMArena, Qwen3.8-Max ranks third on the vision leaderboard at 1,300±8, right behind claude-fable-5 (1,313) and claude-opus-4-7-high (1,301). The real point of that ranking is elsewhere: one secondary source had described the Max version as text-only, and the vision leaderboard settles the question. The 0902 update climbed to first place in Code Arena WebDev at 1,691 – three points ahead of Claude Opus 5 Max and twenty-two ahead of the original 3.8-Max.
On Artificial Analysis, Qwen3.8-Max scores 47 on the Intelligence Index at a blended price of $1.19 per 1M tokens. Gemini 3.8 Flash scores the same 47 at $0.74. One more number worth remembering: 40 output tokens per second, the slowest in the top ten. Gemini 3.8 Flash does 334. Qwen’s flagship is slow, and you’ll feel it in interactive work without looking at any benchmark.
Qwen3.8-Flash-Next – twelve times cheaper than the flagship (August 26, 2026)
An open preview of the Qwen4 architecture. 125B parameters in total, 6B active per token, plus a 51B-parameter layer of N-gram embeddings that lives in ordinary RAM and stays out of GPU memory. Think of it as a dictionary of common word combinations – there’s no need to keep it on the GPU, and that’s where the savings come from.
Context is 262,144 tokens natively, up to 1M via YaRN. Price: about $0.16 input and $0.47 output per 1M tokens, with cached input roughly 89% cheaper. That’s about twelve times less than Qwen3.8-Max.
The model isn’t in our benchmark and, at the time of writing, hadn’t shown up on independent leaderboards either. All we have are vendor numbers: DeepSWE 58.7 and SWE-bench Pro 62.5 on coding, plus CoWorkBench 73.9 and JobBench 55.7. Alibaba also claims Flash-Next beats Qwen3.7-Plus at roughly a ninth of the training cost and outperforms DeepSeek-V4-Flash (284B) and Claude Opus 4.6 on most tested tasks. There’s nowhere to verify that yet, so treat it as a promise.
The weights are on Hugging Face and ModelScope under the Qwen Community License; hosted access is through the QwenCloud API.
Open weights and licensing
Three models from the same generation shipped under three different legal regimes, and for a manager that matters more than the scores.
- Qwen3.8-27B – Apache 2.0. No strings attached: embed it in commercial products, fine-tune it, resell it.
- Qwen3.8-Flash-Next – Qwen Community License. This is Alibaba’s own text, and you need to read it before you deploy.
- Qwen3.8-Max – the first Max-class model promised as open, but it’s reportedly getting a custom license with restrictions for companies earning more than $50 million from AI services. That isn’t Apache 2.0.
Then there’s physics. Qwen3.8-Max in FP8 is about 2.5 TB of weights – data-center territory. For this model, open weights mean you can study it and, with the right hardware and license, deploy it.
A revenue threshold as a license condition isn’t unique: Kimi K3 sets it at $20 million. The practical takeaway: “open-weights model” without the name of the license tells you nothing.
Qwen 3.7 Max – the agentic flagship (May 2026)
Qwen 3.7 Max (also written Qwen3.7-Max) was announced at the Alibaba Cloud Summit on May 20, 2026, after quietly appearing on LM Arena on May 14. The 3.7 series has two models: Qwen3.7-Max (the text flagship for reasoning and agentic work) and Qwen3.7-Plus (the multimodal agent, GA since June 1, 2026).
Qwen3.7-Max key specs:
- 1M-token context – up from 256K in Qwen 3.6 Max
- Agentic architecture – the model can run autonomous sessions of up to 35 hours, chaining 1,000+ consecutive tool calls without losing quality. In the summit demo, it autonomously optimized a kernel – more than a thousand code edits and a 10x inference speedup
- Lowest hallucination rate among flagship models – 22.9%
- “Thinking” mode – a toggleable chain of reasoning, visible in the Qwen Chat interface
Benchmark results:
| Benchmark | Qwen3.7-Max | Comparison |
|---|---|---|
| Intelligence Index v4.0 | 56.6 | 5th in the world, 1st among Chinese models |
| GPQA Diamond | 92.4% | Above Claude Opus 4.6 Max (91.3%) |
| Apex Math | 44.5% | Above Claude Opus 4.6 Max (34.5%) |
| MRCR-v2 128K (long context) | 90.4% | Above Qwen3.6-Plus (85.9%) |
| Terminal-Bench Hard (coding) | 50.8% | – |
Qwen3.7 is API-only: the series has no open weights, and Alibaba hasn’t given any timeline for releasing them. The next generation broke that pattern: 3.8-Max is promised as open, and 27B and Flash-Next shipped with public weights.
Qwen 3.7 Plus – the multimodal agent (GA June 1, 2026)
Qwen3.7-Plus appeared on LM Arena on May 14 as a preview and went GA on June 1, 2026. It’s the multimodal member of the 3.7 series: it takes text, images and video as input and returns text.
How it differs from Qwen3.7-Max:
- GUI Agent Grounding – the model “sees” a screenshot and returns exact pixel coordinates for interface elements. That makes it possible to build autonomous agents that operate applications visually – no API required
- 1M-token context, up to 65,536 output tokens, and an internal reasoning budget of up to 256K tokens
- The
preserve_thinkingparameter – keeps the reasoning state between tool calls in an agent loop - Cost per task – see the card above; among multimodal models, Plus is one of the cheapest options
Qwen 3.7 Plus is a closed model (API-only, no open weights). For a manager, that’s an important signal: the most capable multimodal models in the 3.7 series are cloud-only, with no local deployment option. With Qwen3.8-27B now shipping with vision under Apache 2.0, that argument has lost much of its weight.
Qwen3.6-Plus – the flagship with a 1M-token context
Released in April 2026 as an answer to Gemini 2.5 and GPT-5. A hybrid architecture combines linear attention with a sparse mixture of experts, which lets it handle huge documents without slowing down.
- Up to 1 million tokens of context: an annual report or an entire code repository goes into a single request, no chunking required.
- 78.8% on SWE-bench Verified – close to closed flagships on programming tasks.
- Audio-Visual Vibe Coding: describe the task out loud and hold a mockup up to the camera – the model generates the code.
- Apache 2.0 license – download the weights and run it locally.
Qwen3.5 – a family of multimodal models
Released in March 2026 and available on Microsoft Azure. The key difference is that these models “understand” text, images, audio and video as one whole within a single architecture. A single question can reference an uploaded document, a screenshot, a video clip and text context at the same time.
The series spans the full range, from 0.8B parameters (runs on a smartphone) to 397B (on par with GPT and Claude):
- Qwen3.5-4B / 9B – the “smartest” in their weight classes; they run on a laptop and beat models two to three times their size. Even on a phone.
- Qwen3.5-27B – strict instruction following: structured reports, form filling, multi-step assignments.
- Qwen3.5-35B-A3B – a balance of quality and speed: 35B parameters, only 3B of them active.
- Qwen3.5-122B-A10B – complex analytics and large volumes of information.
- Qwen3.5-397B-A17B – the open-source flagship; on GDPval-AA it beats its predecessor Qwen3-235B by 361 points. Third-best open-source model in the world.
This fall the family is still relevant and still entirely Apache 2.0, with vision on every version up to and including 27B. Hardware rules of thumb: 0.8B runs on a laptop without a GPU and takes about 0.6 GB; 2B fits on an RTX 3060 or an M1; 9B on a 16 GB card; 27B on an RTX 4090. All versions have 256K context natively, and hybrid linear-attention layers remove the usual KV-cache penalty. On machines bound by CPU or memory bandwidth, 35B-A3B runs roughly five times faster than the dense 27B. 122B-A10B and 397B-A17B are text-only and need several H100s.
Common to the whole series:
- About 400 letter-size pages per request; when needed, the context stretches to the equivalent of a short book.
- Two modes – “Thinking” (step-by-step reasoning for hard problems) and “Non-thinking” (instant answers for routine work) – switch within the same model, with no need to change services.
- Support for 201 languages.
- On independent tests, it competes with leading multimodal models at a much lower cost.
Mode switching is the real breakthrough of Qwen3/3.5. You don’t have to pick between a “smart” model and a “fast” one: the same model adapts to the difficulty of the task. Deep reasoning for analyzing a financial report; an instant answer for “when’s the meeting?”.
Qwen3-235B – the open-source text flagship
Released in April 2025 and trained on twice as much data as the previous generation. Its specialized architecture delivers flagship quality at much lower compute cost – a difference you’ll notice when working through the API.
- Open model – any company can download it and run it on its own servers. You’ll need serious server hardware.
- The model “thinks” before answering, checking its logic step by step. That matters for complex calculations, legal analysis and financial planning.
- Qwen3-Max is the cloud version via API, for those who don’t want to deal with server setup.
- Up to 200 letter-size pages per request; up to 400 in extended mode.
Qwen3-30B – the sweet spot for a corporate server
An enterprise-class model: fast and cheap to run despite its size. It outperforms older models that needed three times the resources.
- For teams that want a powerful assistant on their own server without paying for a whole data center.
- Up to 200 letter-size pages per request.
Qwen3-Coder-Next – a model for developers (February 2026)
A specialized model for writing and reviewing code. On independent tests, it beats DeepSeek’s models and holds its own against Claude Sonnet 4.5.
For a manager, this is most likely a tool for the engineering team. The model can hold an entire project in memory, every file in the repository. Running it on your own server takes about 51 GB of RAM.
Qwen 4 Coder – the first open-source coder at 82% SWE-Verified (June 2026)
Qwen 4 Coder 32B-A3B, released June 2, 2026, is the first model to mark the move to Qwen’s fourth generation – and the first open model to break 80% on SWE-Verified, a benchmark that tracks real developer productivity.
- MoE architecture: 32B parameters in total, 3B active per token. 256K context
- 82% SWE-Verified – ahead of Devstral, Qwen3.6-Coder and every model that runs on a laptop
- Runs on a 24 GB Mac: 58 tokens/sec on an M4 Pro – fully usable without a server
- Trained on real tests: unlike its predecessors, Qwen 4 Coder was trained with reinforcement learning on test results from real GitHub issues
- Apache 2.0 license – the same freedoms as Qwen3.8-27B: commercial products, fine-tuning, resale
For a CTO or head of engineering, Qwen 4 Coder is a reason to revisit subscription spend: a model that runs on a developer’s work laptop handles tasks at the level of cloud flagships.
The lineup is impressive, but picking the right model is only part of the job. Alibaba shipped three releases in August, each with its own purpose: a flagship for hard reasoning, a compact model to run in-house and a cheap one for high-volume agent workloads. The gap between “getting an answer” and “getting a reliable result” comes down to how you frame requests and check the output. That’s exactly what our free module is about.
Qwen3.8-Max, a 27B on one GPU and Flash-Next at $0.16 – three releases in a single August. The only way to find out which one is yours is to try them: 9 real manager tasks, free, no sign-up required.
No payment required • Get notified on launch
Qwen3-VL – image and document analysis
Understands images, video (over an hour long), PDFs and handwriting. Text recognition works in 32 languages: upload a contract scan or a photo of the whiteboard after a meeting and get structured text back.

Qwen3.5-Omni – native multimodality (March 2026)
Text, images, audio and video are processed by one architecture, with no separate encoders. By Alibaba’s measurements, it posts best-in-class results on 215 audio and audiovisual understanding tasks. It supports real-time voice input and output, built-in web search and external tool calls. Omni powers the Audio-Visual Vibe Coding workflow: a developer describes the requirements out loud and shows the interface on video, and the model writes the code.
Qwen3-TTS and Qwen-Image-2512 – voice and images
Qwen3-TTS (open license) does speech synthesis with voice cloning and streaming generation for voice assistants and narration. Qwen-Image-2512 is an updated image-generation model with photorealistic detail and high-quality text rendering: multi-line captions, paragraphs, brand elements – exactly where Midjourney and DALL-E fall apart.
Models for local deployment
Alibaba offers a line of compact models you can run on your own hardware, from a laptop to a corporate server:
| Where to run it | Model | Max document | Best for |
|---|---|---|---|
| Smartphone or tablet | Qwen3-0.6B | ~50 pages | Autocomplete, simple suggestions |
| Standard laptop (4 GB RAM) | Qwen3-1.7B | ~50 pages | Email drafts, short summaries |
| Powerful laptop (8 GB RAM) | Qwen3-4B | ~50 pages | Document analysis, quality on par with older top models |
| Workstation (16 GB RAM) | Qwen3-8B | ~200 pages | Complex documents, spreadsheets, long conversations |
| Department server (32 GB RAM) | Qwen3-14B | ~200 pages | A full-fledged team assistant |
| Corporate server (48 GB RAM) | Qwen3-32B | ~200 pages | Quality close to top cloud models |
| Server cluster (140 GB+) | Qwen3-235B | ~400 pages | Flagship quality, competes with GPT and Claude |
If your company can’t send data to external servers – finance, healthcare, the public sector – you can download the model you need and deploy it inside your corporate network. The data never leaves your perimeter. You can’t do that with ChatGPT or Claude.
Qwen3-4B fits on an ordinary laptop (4 GB of memory and up) and delivers quality comparable to models that used to need a full server rack.
A hands-on guide to choosing hardware, installing Ollama and comparing local models with cloud ones is in our local LLM guide for managers.
Qwen3.8-27B – a model that fits on a single GPU
Released August 13–14, 2026, and for a company this matters more than the flagship launch.
A dense 27–28B model with a vision encoder: it natively understands images and video and returns text. 262K context. Apache 2.0 – no revenue caveats and no separate legal review for every use case. At Q4_K_M quantization it’s about 17 GB, meaning one consumer GPU in the RTX 4090 class or a laptop with 16 GB of memory via Ollama or llama.cpp.
It isn’t in our benchmark yet, but there’s little to compare it against: Apache 2.0, vision out of the box and a single GPU – no model on the market offered that combination before August.
For a company that isn’t allowed to send data to the cloud, this is the first “everything in-house” option that doesn’t feel like a compromise. Qwen3-4B on a laptop was a proof of concept. A 27B with vision and 262K context is a working tool for contract scans, whiteboard photos, internal policies and correspondence you can’t show an outside service.
What to check before you roll it out:
- Speed on your hardware. 17 GB is the size of the weights; it says nothing about acceptable response time. On a machine with 16 GB of memory, part of the model spills over to the CPU and generation slows down.
- Quality on your own documents. There are no independent measurements yet, so test it on your real material – a polished demo proves nothing here.
- Maintenance. Local deployment means a server, weight updates, monitoring and a person who owns all of it. Part of your API savings will go into that person’s salary.
How Qwen stacks up against the competition
Independent leaderboards paint a more modest picture than the launch announcements. In the September 2026 Artificial Analysis index, the gap between first place (Claude Fable 5.1, 57 points) and tenth (DeepSeek V4 Pro 0813, 42) is 15 points, while the price gap is 34x: $6.12 versus $0.18 for GLM-5.3-Flash. Qwen3.8-Max’s 47 points are shared with Gemini 3.8 Flash.
A word on methodology. Artificial Analysis publishes its index in versions and rescales between them, so the number in a model’s launch post and the number in the live table sit on different scales. You can only compare within a single snapshot.
On LMArena, positions one through nine overlap within their confidence intervals on both leaderboards, text and vision. Formally, it’s a tie. Anyone crowning a best model from a table like that is reading it too literally – we have a separate breakdown on why.
| Criterion | Qwen | ChatGPT / Claude | DeepSeek |
|---|---|---|---|
| Flagship quality | Top tier (47 on Artificial Analysis, #1 in Code Arena WebDev) | Top tier | Mid-range at a rock-bottom price |
| Free web access | Yes, chat.qwen.ai with no limits | Limited | Limited |
| Can run on your own server | Yes – Qwen3.8-27B, Flash-Next, Qwen 4 Coder, the 3.6 and 3.5 series; the 3.7 series is closed | No | Partially |
| API cost | Lowest in class: Qwen3.8-Max – $2.00/$6.00, Flash-Next – $0.16/$0.47 | High: GPT-6 Astra – $10/$50 | Low |
| Image and video | Yes (3.8-Max, 3.8-27B, 3.7 Plus, 3.5-Omni) | Yes | Limited: vision only in preview on the cheap tier |
For most business tasks – writing, data analysis, document summaries – Qwen 3.7 Plus delivers results comparable to Claude Sonnet 5 at a noticeably lower price.
Qwen3.8-Max costs $2/$6 per million tokens where GPT-6 Astra asks $10/$50 – and neither number means anything until you test it on your own task. The free module has 9 hands-on manager tasks, no sign-up required.
No payment required • Get notified on launch
Results in our benchmark (updated September 2026)
Qwen is the most widely represented ecosystem in our benchmark: five models. Results are based on 80 scenarios drawn from real manager work (analysis, planning, communication, team management). The 3.8 lineup isn’t among them: we haven’t tested Qwen3.8-Max, 27B or Flash-Next yet, and everything said about them above relies on vendor data and independent leaderboards.
Three flagships – Qwen 3.6 Plus, Qwen 3.7 Max and Qwen 3.7 Plus – are covered above with fresh benchmark data (rank, score, cost per task and test date are in the cards under their headings). Here are the two remaining mid-tier models:
Qwen 3.6 Flash beats Gemini 3.5 Flash at a fraction of the price; Qwen 3.6 27B is especially strong at planning. On price-to-quality, Qwen 3.6 Plus looks more compelling than both Claude Sonnet 5 and GPT-5.4 Mini, and it costs less.
The lineup’s main paradox: Qwen 3.7 Max costs more than Qwen 3.6 Plus but ranks lower overall. Where it wins is specifically “planning” and “team management” – if those are your core tasks, Max is worth it.
Qwen 3.6 Plus’s profile across eight categories makes the gap obvious: strong on text and analysis, weak wherever local context is required.
The whole Qwen lineup shares the same weaknesses:
- Local employment law – on our benchmark’s scenarios, built around Russian labor law, it invents nonexistent statutes and penalties. Don’t settle critical legal questions without checking.
- Financial calculations – systematic arithmetic errors.
- Regional awareness – the weakest category for every Qwen model.
Here’s the usage strategy for now. Qwen 3.6 Plus is the main workhorse. Qwen 3.7 Plus handles high-volume routine work when budget is tight. Max is justified where planning and team-management tasks are what matter: by our measurements, it’s stronger in those categories. Once the 3.8 lineup is in the benchmark, this picture will need to be recalculated.
Try it yourself: planning + employment law
Qwen usually produces an excellent plan structure (planning is its strongest category), but when it cites specific employment rules it often invents statutes and penalties that don’t exist (regional awareness is its weakest category). DeepSeek has the same problem with local law, and Kimi K2.6 is the strongest planner but also knows local regulations poorly.
Practical uses for managers
Most managers who try Qwen start with Qwen Chat, a free web interface with access to current models. It supports web search, image generation and document analysis with no subscription. It’s a good way to evaluate the tool before you talk to IT.
Qwen handles messy documents with confidence: contract scans, handwritten notes and whiteboard photos come back as clean, structured text.
A separate use case is enterprise deployment without the cloud. For companies that can’t send data outside their perimeter, Qwen offers something ChatGPT and Claude don’t: you can download the models and run them on your own servers. Since August 2026, that no longer means compromising on quality – Qwen3.8-27B with vision and Apache 2.0 covers it on a single GPU. It’s the same logic governments follow when building sovereign AI systems: control over data beats the convenience of the cloud.
Finally, Qwen3.5-Omni and Qwen3-TTS let you build a voice assistant end to end – speech recognition, request handling and speech synthesis all from one ecosystem, with no need to stitch several services together.
The U-turn: the flagship goes open again
In the first half of 2026, Qwen’s core value proposition shifted. Alibaba moved its flagships – Qwen3.7-Max and Qwen3.7-Plus – to a closed format (API-only, no open weights). That put Alibaba in direct competition with OpenAI and Anthropic for enterprise API contracts and looked like a retreat from the very thing the ecosystem was built on.
Even then it wasn’t a full exit from open source: Qwen3.6-35B-A3B and Qwen 4 Coder shipped under Apache 2.0. But the most capable models were closed, and the company made no firm commitment on when the 3.7 series might open up.
In August, the direction changed. Qwen3.8-Max is the first Max-class flagship whose weights Alibaba has promised to publish. The same month brought Qwen3.8-27B under Apache 2.0 and Qwen3.8-Flash-Next with weights on Hugging Face and ModelScope. The 3.7 series stayed closed.
The spring pivot coincided with key people leaving. In March 2026, technical lead Lin Junyang stepped down, posting: “me stepping down. bye my beloved qwen”. The same day Yu Bowen, who ran post-training, resigned. Hui Binyuan, the lead code researcher, had left earlier. A teammate commented: “I know leaving wasn’t your choice.” Alibaba’s stock fell 5.3% that day.
For a manager choosing Qwen, here’s the picture now:
- Open models have reached the top tier. 27B and Flash-Next are already downloadable, and Max is promised. If data control is critical, there’s nothing left to wait for.
- You have to read the terms. Apache 2.0 for 27B, a custom license for Flash-Next, a revenue threshold for Max.
- The staffing instability hasn’t gone away. The release pace didn’t drop over the summer, but one season is too short to call a trend.
Limitations and risks
- Flagship license: Qwen3.8-Max has been promised as open, but it will come with a custom license and a revenue threshold. Only plan a commercial on-premises deployment after you’ve read the text.
- Flagship size: about 2.5 TB of weights in FP8, data-center scale; in practice Qwen3.8-Max stays a cloud model whether or not the weights are published.
- Speed: 40 output tokens per second according to Artificial Analysis – the slowest in the top ten. That’s noticeable in interactive work and in agent loops with hundreds of calls.
- Staff losses: the departure of the project’s technical lead and its head of post-training in March 2026 is a serious signal. Alibaba reorganized the team into a cross-group foundation-model unit, but the long-term consequences are still unclear.
- Censorship and filters: like other Chinese models, Qwen has built-in restrictions on certain political and historical topics. For business tasks this rarely becomes a problem.
- Geopolitical risk: the model belongs to Alibaba (China). For some organizations that can rule out the cloud API. The workaround is local deployment of the open versions.
- Quality outside English: despite support for 200+ languages, answers in other languages can lag behind English ones. Test on your own tasks.
- Integrations: there are far fewer ready-made connectors to Western services – Slack, Notion, Google Workspace – than for ChatGPT or Claude. Qwen is better integrated with Chinese enterprise tools.
Pricing and availability
| Option | Cost | Best for |
|---|---|---|
| chat.qwen.ai | Free | Trying Qwen without an API account |
| Qwen3.8-27B (download) | Free, Apache 2.0 | Vision, 262K context, ~17 GB at Q4_K_M – one GPU |
| Qwen3.8-Flash-Next (API / weights) | $0.16 / $0.47 per 1M input/output | High-volume agent workloads, up to 1M context via YaRN |
| Qwen3.8-Max (API) | $2.00 / $6.00 per 1M, cache $0.25 | Flagship: text, images and video, 1M context |
| Qwen 4 Coder 32B-A3B (download) | Free (Apache 2.0) | Coding on a laptop, 82% SWE-Verified |
| Qwen3-0.6B / 4B (download) | Free | Running on a laptop, full control over data |
| Qwen3-235B (download) | Free | Flagship locally, needs a server cluster |
| Qwen3.7-Plus (API, GA) | $0.40 / $1.60 per 1M input/output | Multimodal agent, GUI grounding, 1M context |
| Qwen3.7-Max (API) | $2.50 / $7.50 per 1M input/output | Text flagship, agentic workloads, 1M context |
| Qwen3.5-397B-A17B (API) | $0.39 / $2.34 per 1M input/output | Open-license flagship |
For comparison: analyzing a 100-page report with GPT-4o costs about $0.35, with Claude Sonnet about $0.55. At the top end the gap is wider: $10/$50 per million tokens for GPT-6 Astra versus $2/$6 for Qwen3.8-Max.
Qwen still offers three ways in: a free chat, a cheap cloud API and local deployment. Over the summer, the third one changed: a 27B with vision and a promised open Max put local deployment in the top tier.
Is it worth trying?
Six months ago, the case for Qwen was simple: “everything ChatGPT does, but free and open.” By spring it was wobbling. By September it was back – with a stronger lineup.
Open models now cover the full range of tasks. Qwen3.8-27B with Apache 2.0, vision and 17 GB of weights runs on a single GPU. Flash-Next at $0.16/$0.47 handles agent workloads with up to 1M tokens of context, and Qwen 4 Coder and the Qwen 3.5 family are still going strong. For companies that need local control over their data, there still aren’t many alternatives, and this is the strongest one available.
On most tasks, all the top models deliver comparable results today, and the difference isn’t obvious in day-to-day work. Qwen 3.7 Plus covers multimodal analysis, agent workflows and screen work – the same ground as Claude Sonnet 5, at several times lower cost. For high-volume work, that’s real money.
The departure of key developers and the team reorganization are a risk worth watching. The release pace hasn’t slowed so far, but that’s no guarantee. If you’re building critical processes on Qwen, keep a fallback – DeepSeek or cloud models backed by a more stable corporate structure.
This fall, keep an eye on one thing: the text of the Qwen3.8-Max license. The promise to open the flagship has been made; the terms you’ll be able to use it on, nobody has seen yet.
You have the tool. Now build the skill
The foundation – prompt engineering, critical thinking, ethics and security – plus role-based tracks: people management, project management, education. The exercises aren't tied to one tool and work in Qwen, DeepSeek, ChatGPT and Claude. Nine hands-on manager tasks are free, no sign-up required.
Sources
- Qwen3.8-Max 0902 on OpenRouter – model card: modalities, context, pricing
- Alibaba releases Qwen3.8-Flash-Next – Flash-Next architecture and pricing
- Qwen3.8-Max: DataCamp breakdown – benchmarks and license terms
- Alibaba releases Qwen3.8-Max – specs and release timeline
- Qwen 3.5: the complete guide – hardware requirements for the open family
- Artificial Analysis – model leaderboard – Intelligence Index, blended price, output speed
- LMArena (Arena.ai) – preference leaderboard – text and vision leaderboards, September 2026
- Qwen Chat – free web interface – Alibaba’s official chat, works without registration
- Official Qwen website – documentation and descriptions of all models
- Qwen models on Hugging Face – 433 open models to download, 300M+ downloads
- Qwen on GitHub – source code and documentation, 40+ repositories
- API via Alibaba Cloud Model Studio – cloud access to Qwen-Plus and Qwen-Max

Stanislav Belyaev
Engineering Leader at Microsoft18 years leading engineering teams. Founder of mysummit.school. 700+ graduates at Yandex Practicum and Stratoplan.

















