As AI consultants based in Rosenheim, Germany, we recommend Qwen models for enterprises looking for an open-weight alternative. With the Qwen 3.8 generation (August 2026), Alibaba ships three very different building blocks: the compact, Apache 2.0-licensed Qwen3.8-27B for local deployment, the MoE flagship Qwen3.8-Max with 2.4 trillion parameters – and, since 26 August 2026, Qwen3.8-Flash-Next, an open preview of the architecture behind the upcoming Qwen4 generation. For GDPR-compliant usage in the DACH region (Germany, Austria, Switzerland), we recommend self-hosting on EU infrastructure. This overview reflects the state as of October 2026.
Qwen 3.8: Models and Licences at a Glance
The Qwen 3.8 generation draws a clear line between a freely usable workhorse, a licence-fenced flagship and an experimental architecture preview with its own community licence.
Qwen3.8-Flash-Next – open preview of the Qwen4 architecture
On 26 August 2026, the Qwen team published Qwen3.8-Flash-Next on Hugging Face and ModelScope. Per the model card, it is an “experimental preview of the architecture that will underpin Qwen4” – explicitly not a GA model, but a look at the next generation. The key facts per the model card:
- 125 billion parameters, only 6 billion active per token, complemented by an n-gram embedding with 51 billion parameters and a multi-token-prediction module with 4 billion parameters
- 262,144 tokens native context, extensible to 1 million tokens via YaRN
- Multimodal: text, image and video input
- Licence:
qwen-community-1.0– not Apache 2.0, even though some press reports say so. The licence text in the repository is what counts. - Vendor-reported: SWE-bench Pro 62.5 · LiveCodeBench v6 91.9 · CoWorkBench 73.9 · GPQA Diamond 91.7 · MathVision 95.7 · AndroidWorld 84.5
- A separate FP8 checkpoint (Qwen3.8-Flash-Next-FP8) is available under the same licence; per Qwen, performance is nearly identical to the BF16 original
Independently measured, Qwen3.8-Flash-Next scores 56 points on the Artificial Analysis Intelligence Index v4.1.1 at around 86 tokens/s output speed – but produced around 200 million output tokens for the index. As always: measure token consumption against your own workload.
In parallel, Alibaba released Qwen3.8-Flash as a production API via Alibaba Cloud Model Studio (model ID qwen3.8-flash). Per Alibaba, the API natively supports a 1 million token context and is compatible with the OpenAI and Anthropic protocols; per secondary sources the output limit is up to 131k tokens. Alibaba’s price list shows US$0.15 / 1M input and US$0.47 / 1M output (Singapore and Frankfurt with scope “EU”), and US$0.113 and US$0.382 respectively in the “Global” scope. Since 27 October 2026, Qwen3.8-Flash has also been available in Frankfurt with the “EU” deployment scope (see the GDPR section).
Context: For self-hosting, Qwen3.8-Flash-Next is very throughput-friendly thanks to its 6 billion active parameters, but needs memory for the full roughly 180 billion parameters including the embedding table. If you need a production-ready, straightforwardly licensed base today, stay with Qwen3.8-27B (Apache 2.0). If you want to evaluate the Qwen4 architecture early, Flash-Next gives you an open candidate. For Qwen 4 itself, Alibaba has not yet published a release date, model sizes or pricing.
Qwen3.8-27B – our new open-weight top pick
On 14 August 2026, the Qwen team published the weights of Qwen3.8-27B on Hugging Face and ModelScope – a dense model with roughly 27.8 billion parameters under Apache 2.0:
- 262,144 tokens native context, extensible to 1 million tokens via YaRN RoPE scaling (vLLM, SGLang)
- Native image and video processing – from STEM diagrams and documents to hour-scale videos
- Thinking mode on by default, but disableable per request; reasoning depth is controlled via the
reasoning_effortparameter - Tool calling through the common inference frameworks (vLLM, SGLang, Ollama, llama.cpp, LM Studio)
- Vendor-reported benchmarks: SWE-Bench Pro 61.7 · GPQA Diamond 89.2 · OSWorld-Verified 84.3 · MathVision 94.6
The decisive practical point is size: quantised, the model fits into the memory of a 24 GB card (e.g. RTX 3090) or a suitably equipped laptop. That makes a capable multimodal model locally operable – without any data leaving the company.
Context: Almost all launch benchmark figures came from the vendor. Independently measured: Artificial Analysis places Qwen3.8-27B in
xhighmode at 52 points on the Intelligence Index, listing it as the top-ranked open-weight model on its list. The price is compute: for the index the model produced roughly 160 million output tokens (median for comparable models: around 45 million) at a measured output speed of about 54 tokens/s. That matches the recurring hands-on finding from the community: in its default setting the model “thinks” considerably longer than necessary – if latency and token cost matter, configurereasoning_effortexplicitly.
Qwen3.8-Max – flagship with a bespoke licence
Qwen3.8-Max has been available via Alibaba Cloud Model Studio since 2 August 2026: a Mixture-of-Experts model with 2.4 trillion total parameters and 95 billion active parameters.
- 1 million token context window, max. 991,808 tokens input (983,616 in thinking mode), up to 131,072 tokens output, chain-of-thought budget 262,144 tokens
- Text, image and video input, text output
- Function calling and structured output are supported; batch inference and fine-tuning are not
- Regions: China (Beijing), Singapore, Germany (Frankfurt), Japan (Tokyo), Hong Kong, US (Virginia); in Frankfurt with deployment scope “Global” only
- Alibaba list prices: from US$1.65 / 1M input and US$4.951 / 1M output; US$2.00 and US$6.00 respectively in the Singapore region
- Since 2 September 2026 there has also been the snapshot
qwen3.8-max-0902(aliasqwen3.8-max-2026-09-02), with improved coding and agent behaviour per Alibaba, at the same price
Roughly ten days after the API launch, Alibaba also released the weights as Qwen3.8-2.4T-A95B – a first for a Max-class model. Two constraints matter here:
- The open weights are a text-only model. Multimodal input (image, video) exists only in the API variant. In addition, thinking mode cannot be disabled in the open-weight version.
- Not Apache 2.0. Instead of the previous permissive licence, a bespoke Qwen3.8-Max licence applies. According to reports it requires a separate commercial licence from companies that use the model in a “model as a service” or “AI work assistant” business and generate more than US$50 million in revenue over any consecutive twelve-month period. Purely internal use without third-party access is reported to be exempt.
In practice: For most mid-sized users in the DACH region the revenue threshold is not relevant. However, anyone planning to resell Qwen3.8-2.4T-A95B as part of their own AI product should have the licence text reviewed legally before rollout – experience with Apache 2.0 from earlier Qwen generations does not carry over here.
Independent evaluation vs. vendor figures
There are noticeable differences between vendor and third-party measurements:
| Source | Statement on Qwen3.8-Max |
|---|---|
| Alibaba (vendor-reported) | Terminal-Bench 2.1 86.6 · SWE-Bench Pro 67.7 · OSWorld-Verified 86.1 · GPQA Diamond 92.6 |
| Artificial Analysis (independent) | 58 points on the Intelligence Index, rank 10 of 188 models; at the same time low output speed (~45 tokens/s) and very high token consumption |
An independent measurement is now available for Qwen3.8-27B as well (since 17 August 2026): 52 points on the Artificial Analysis Intelligence Index (xhigh mode). The size comparison is notable – the model reaches that score with roughly 27.8 billion parameters, while comparably rated open models are considerably larger. The score is bought with very long reasoning chains, though (see the note above).
Notably, the Artificial Analysis figure was revised upward several times between the first measurement and later re-runs (53 → 56 → 58). Such variation is a good reason not to treat benchmark leaderboards as the sole basis for selection, but to test against your own tasks. On LMArena, media reports placed Qwen3.8-Max as the highest-ranked Chinese model for text tasks in early August.
Qwen-Image-2.1 and Qwen-Image-2.1-Turbo – image generation and editing
Qwen-Image-2.1 (20 September 2026) is a text-to-image and image-editing model with a 7-billion-parameter generator, native transparency (RGBA) and up to 10 reference images. Per its model card, Qwen-Image-2.1-Turbo is an accelerated checkpoint of the same model: it generates images in 8 denoising steps, uses the same 7B architecture, ships with its recommended sampling schedule (default CFG=1) and uses prefix KV caching to reuse text and reference-image context across steps. It loads directly with the QwenImage21Pipeline in Diffusers and appeared on Hugging Face in October 2026 (Qwen/Qwen-Image-2.1-Turbo). Both checkpoints are under the Qwen Research License, which in our reading means non-commercial use only; commercial use requires a separate agreement with Alibaba. The model card gives no benchmark figures.
Qwen3.8-Omni-Flash – native omnimodal model
Per the Alibaba Model Studio documentation, Qwen3.8-Omni-Flash (model ID qwen3.8-omni-flash, 17 September 2026) is a native omnimodal model: it processes text, image, audio and video as input with text-only output, offers 1 million tokens of context, function calling and web search, supports audio in 113 languages and is reachable via the Chat Completions and Responses interfaces. A realtime variant (qwen3.8-omni-flash-realtime) has been listed since 21 September 2026. In Frankfurt the model can only be booked with deployment scope “Global”. No open weights were announced. Prices are in the Pricing section.
Qwen 3.7: The Previous Flagship Generation
At the Apsara Summit in Hangzhou on 20 May 2026, Alibaba unveiled Qwen3.7-Max – the new proprietary top model of the Qwen family. On 1 June 2026, Qwen3.7-Plus followed as the multimodal agent variant at roughly one-sixth of the per-token cost.
Qwen3.7-Max
- 1 million token context window – fits large-scale agent workflows with long documents
- Top scores (per Alibaba) on Terminal-Bench 2.0, SWE-Bench Pro and MCP-Atlas
- Support for the Anthropic protocol (MCP)
- Available via Alibaba Cloud Model Studio (proprietary, no open weights)
Qwen3.7-Plus
- Multimodal agent (text, image, video) on the Qwen 3.7 backbone
- Deep reasoning, tool invocation and autonomous iteration via the Bailian platform
- Listed at roughly 1/6 of Qwen3.7-Max pricing – attractive for vision-heavy agent workloads
- First seen as preview on LM Arena (around 14 May 2026), GA on 1 June 2026
Note: The Qwen 3.7 line remains API-only via Alibaba Cloud. For self-hosting we have recommended Qwen3.8-27B (Apache 2.0) since August 2026.
Qwen 3.6: Agentic Real-World AI
In April 2026, Alibaba released the Qwen 3.6 generation – focused on real-world agent workflows:
Qwen3.6-35B-A3B – still the most frugal option
The model uses a MoE architecture (36B total, only 3B active) and, per Alibaba, outperforms Google Gemma 4 in benchmarks. The Apache 2.0 license enables unrestricted self-hosting. As a general open-weight recommendation it has been superseded by Qwen3.8-27B, but it remains the choice when throughput per GPU hour outweighs model quality.
Qwen3.6-27B – Dense Alternative
With 28 billion parameters, this dense model delivers strong performance for applications that don’t support MoE architecture. It has since been superseded by Qwen3.8-27B; STACKIT lists Qwen3.6-27B as deprecated.
Qwen3.6-Plus and Max-Preview
The proprietary models are only available via API and reflect Alibaba’s strategic shift toward commercial offerings. Qwen3.6-Plus focuses on real-world agents – autonomous AI that executes real tasks like app control and document editing. Alibaba removed Qwen3.6-Max-Preview from Model Studio on 10 October 2026 (successor: Qwen3.7-Max).
Note: With Qwen 3.6, Alibaba introduces proprietary models for the first time that are not available as open source. For self-hosting we now recommend Qwen3.8-27B; Qwen3.6-35B-A3B (Apache 2.0) remains the option for maximum throughput per GPU.
Qwen Agent & agentic Qwen workflows
Beyond the models themselves, the Qwen team maintains Qwen-Agent (github.com/QwenLM/Qwen-Agent), an open-source framework for building AI agents on top of the Qwen models (Qwen 3.0 and newer). It encapsulates the typical building blocks of production agents and removes most of the boilerplate for developers:
- Function calling / tool use: native support for parallel, multi-step, and multi-turn tool calls, including automatic parsing and execution
- MCP integration: connect external tools and data sources (e.g. GitHub, filesystem, SQLite) via the open Model Context Protocol (MCP) standard
- Code interpreter: execute Python code in an isolated Docker sandbox
- RAG over long contexts: document QA over 1M+ tokens using hybrid RAG and agent-based decomposition
- BrowserQwen: a browser assistant with a Chrome extension as a reference implementation
The framework supports Qwen3, Qwen3-VL, Qwen3-Omni, Qwen3-Coder and QwQ, among others, and can be run either via the Alibaba Cloud API or against local, OpenAI-compatible servers (vLLM, SGLang).
On the model side, Alibaba is pushing the agentic direction hard. Qwen3.7-Max is explicitly positioned as the “Agent Frontier”: Alibaba demonstrates (as of June 2026) long autonomous runs – for instance a 35-hour kernel-optimization run with over 1,000 tool calls – and reports top scores on tool-use benchmarks such as MCP-Mark and Terminal-Bench 2.0. According to the vendor, the model serves as a drop-in backbone for agent frameworks (including Claude Code, Qwen Code, OpenClaw) and supports MCP. On the open-weight side, Qwen3.6-35B-A3B targets agentic coding explicitly and works, for example, with the open-source terminal agent OpenClaw.
Sovereignty & GDPR: The cleanest path for agentic Qwen workflows in the DACH region is the combination of open Qwen weights (e.g. Qwen3.8-27B, Apache 2.0) plus Qwen-Agent, fully self-hosted on EU infrastructure – so neither prompts nor tool data leave your company. Details in the next section.
GDPR: What to Consider When Using Qwen
Qwen is a model from a Chinese vendor. For the data-protection assessment, what matters is therefore less the model itself than the question of where it is operated and where data flows in the process.
Self-hosting in the EU – the most straightforward route
For Qwen3.8-27B and the other Apache 2.0 models: the weights can be downloaded and run entirely on your own or on European infrastructure (vLLM, SGLang, Ollama, llama.cpp). No transfer to Alibaba occurs; processing takes place solely within your own sphere of responsibility. Prompts, documents and tool calls stay in-house – the core requirement of many data protection officers.
Besides your own hardware, European inference providers that run open models in EU data centres are an option – for example IONOS AI Model Hub, STACKIT, Scaleway or Nebius. Which Qwen version each of them carries changes quickly; this should be verified concretely before deciding.
Using the Alibaba Cloud API
Alibaba Cloud Model Studio operates a Germany (Frankfurt) region with the deployment scopes “Global” and “EU”. Per Alibaba, the region determines where request data is stored and the scope determines where inference runs. The “EU” scope restricts inference to the European Union, but it only applies to selected models: per the price list, these include qwen3.8-flash (since 27 October 2026), qwen3.5-flash, qwen-plus and qwen3-vl-plus. Qwen3.8-Max and Qwen3.8-Omni-Flash remain available in Frankfurt with scope “Global” only: data then rests in Frankfurt, but inference may take place worldwide. Within the Qwen 3.8 line, Qwen3.8-Flash is therefore the only model that can be booked via Model Studio with EU-restricted inference. Individual add-on features such as built-in web search are also unavailable in Frankfurt.
To put this factually: there is no adequacy decision by the European Commission for the People’s Republic of China. Transferring personal data there is therefore only permissible under the conditions of Art. 44 et seq. GDPR – in practice usually via standard contractual clauses (Art. 46(2)(c) GDPR) together with a transfer impact assessment. Potential access powers under Chinese law must be considered, as must the question of whether support, operations or administration access can occur from a third country. Whether an EU deployment scope fully resolves these questions in a specific case depends on the contractual arrangement and must be assessed individually.
This is a general technical assessment and does not replace a legal review of your specific use case. For the contractual and regulatory evaluation, please involve your legal advisors – we are happy to support the technical part as part of our AI compliance consulting.
Managed alternatives with EU data residency
Qwen3 models are available as fully managed in-region endpoints via AWS Bedrock – per the current Bedrock model cards, for example Qwen3-32B in Frankfurt (eu-central-1), Ireland (eu-west-1), London (eu-west-2, UK), Stockholm (eu-north-1) and Milan (eu-south-1), plus Qwen3-235B-A22B-2507, Qwen3-Next-80B-A3B, Qwen3-Coder-Next and others, with regions varying by model and endpoint. Qwen3-Coder-30B-A3B-Instruct and Qwen3-Coder-480B-A35B-Instruct have been Legacy there since 30 September 2026, with end of life on 30 March 2027. The Qwen 3.8 generation is not listed on Bedrock.
For the current models with EU data residency this leaves three routes: self-hosting Qwen3.8-27B, STACKIT AI Model Serving (Qwen3.8-27B as a managed API in region eu01) and Qwen3.8-Flash via Model Studio Frankfurt with scope “EU” – the latter with the questions about the provider and third-country access described above.
QwenWork: agent platform without an EU route for now
Alongside the 3.8 generation, Alibaba opened the public beta of QwenWork on 3 August 2026 – an agent platform for workplace workflows that bundles desktop, cloud and collaboration agents on top of Qwen 3.8. Per Alibaba the beta is initially available in China; embedding into DingTalk is announced. For DACH enterprises QwenWork is therefore not a practical sourcing option at present – the relevant routes remain open weights or Model Studio.
Qwen 3.5: Multimodality and Efficiency
Native Multimodality
Qwen 3.5 unifies text, image, and video in one architecture:
- Video Analysis: Understands up to 2 hours of video in a single prompt
- Timestamp-Precise: Identifies events at second-level resolution
- Long Context: Up to 1 million tokens (entire books, large codebases)
- Flexible Input: URLs, local files, frame sequences
Agentic AI & Automation
Qwen 3.5 can execute autonomous workflows:
- App interaction on smartphones
- Document editing and email management
- Travel booking and process automation
- Multi-step tasks with tool use
Efficiency through Mixture-of-Experts
The Qwen3.5-397B-A17B model uses:
- 397 billion parameters total
- Only 17 billion active per inference
- 60% lower costs than predecessors
- 8-19x higher throughput than Qwen3
Qwen 3.5 Small Model Series
In March 2026, Alibaba released a new series of compact models for edge and mobile applications:
| Model | Parameters | Use Case |
|---|---|---|
| Qwen3.5-9B | 9B | Edge servers, rivals 30B+ models |
| Qwen3.5-4B | 4B | UI navigation, document analysis |
| Qwen3.5-2B | 2B | Mobile devices |
| Qwen3.5-0.8B | 0.8B | IoT, smartphones |
All small models are natively multimodal and agent-capable. They are particularly suited for on-device AI where privacy is ensured through local processing.
Qwen3.5-Max-Preview
At launch, per Alibaba, Qwen3.5-Max-Preview ranked ahead of the other Chinese models on LM Arena and reached 5th place globally in math reasoning. The preview build has since been superseded by the GA models of the 3.7 and 3.8 lines.
Key Strengths
Open Weights – but Check the Licence
- Full Control: Model runs in your infrastructure
- No Licence Costs for the Apache 2.0 models (incl. Qwen3.8-27B, Qwen3.6-35B-A3B, Qwen3-Coder), commercial use permitted
- Customizable: Fine-tuning on your own data possible
- GDPR-Friendly: No data leaves your company
- Exception: The open weights of Qwen3.8-Max carry a bespoke licence with a revenue threshold for MaaS providers – a separate review is required here
Expanded Multilingual Support
Qwen 3.5 supports 200+ languages and dialects:
- Chinese (outstanding)
- European languages (very good)
- Expanded coverage: South Asia, Africa, Oceania
- Competitive with Western models
Text-in-Image Generation
Per Alibaba, Qwen-Image is designed for:
- Complex text layouts
- Multilingual text rendering
- Paragraph-level semantics
- Fine detail work
Availability
| Route | Models | Note |
|---|---|---|
| Hugging Face / ModelScope | Qwen3.8-27B (Apache 2.0), Qwen3.8-Flash-Next (qwen-community-1.0), Qwen3.8-2.4T-A95B (bespoke licence), earlier Qwen generations | Weight downloads incl. FP8 variants |
| Self-hosting | vLLM, SGLang, Ollama, llama.cpp, LM Studio | 262k context natively, up to 1M via YaRN |
| Alibaba Cloud Model Studio | Qwen3.8-Max, Qwen3.8-Flash, Qwen3.8-Omni-Flash, Qwen 3.7/3.6/3.5 lines | Regions: China (Beijing), Singapore, Germany (Frankfurt), Japan (Tokyo), Hong Kong, US (Virginia); in Frankfurt scope “EU” for Qwen3.8-Flash and others, otherwise “Global”; OpenAI-, Anthropic- and DashScope-compatible endpoints; check regional availability per model |
| AWS Bedrock | Qwen3-32B, Qwen3-235B-A22B-2507, Qwen3-Next-80B-A3B, Qwen3-Coder-Next and others | Managed, serverless, in-region – depending on the model incl. Frankfurt (eu-central-1), Ireland (eu-west-1), London (eu-west-2, UK), Stockholm (eu-north-1), Milan (eu-south-1); Qwen3-Coder-30B/-480B Legacy (end of life 30 March 2027); Qwen 3.8 not listed there |
| Microsoft Foundry | Qwen3.8-27B and Qwen3.6-35B-A3B (preview), Qwen3-32B | Per Microsoft listed as fine-tuning models only, for Qwen3.8-27B with “Global” training and no data zone; no serverless offering for the Qwen 3.8 line |
| Google Cloud (Gemini Enterprise Agent Platform, formerly Vertex AI) | Qwen3 models (self-deployment) | Per Google, all Qwen managed (MaaS) endpoints were shut down on 21 October 2026 – only self-deployment via Model Garden, no region guarantee |
| OpenRouter | Qwen3.8-Max, Qwen3.8-27B and others | Routed access via rotating providers |
| STACKIT AI Model Serving | Qwen3.8-27B, Qwen3-VL-235B-A22B-Instruct (FP8), Qwen3-VL-Embedding-8B; Qwen3.6-27B deprecated per STACKIT | Managed, OpenAI-compatible API in region eu01 (Germany) per the STACKIT documentation |
| Other EU inference providers | open Qwen models | e.g. IONOS AI Model Hub, Scaleway, Nebius – check catalogue first |
Pricing
| Model | Input / 1M | Output / 1M | Source |
|---|---|---|---|
| Qwen3.8-Max / -0902 (Global scope, incl. Frankfurt) | $1.65 | $4.951 | Alibaba Cloud Model Studio |
| Qwen3.8-Max / -0902 (Singapore) | $2.00 | $6.00 | Alibaba Cloud Model Studio |
| Qwen3.8-Flash (Singapore and Frankfurt with EU scope / Global scope) | $0.15 / $0.113 | $0.47 / $0.382 | Alibaba Cloud Model Studio |
| Qwen3.8-Omni-Flash (Singapore / Global scope) | $0.15 / $0.113 | $0.47 / $0.382 | Alibaba Cloud Model Studio |
| Qwen3.8-27B (API, Singapore) | $0.50 | $3.00 | Alibaba Cloud Model Studio |
| Qwen3.8-27B / Qwen3.8-Flash-Next self-hosted | — | — | no licence fees (Apache 2.0 or qwen-community-1.0), infrastructure only |
Prices per Alibaba, as of October 2026.
Benchmarks & Performance
| Benchmark | Score | Model | Origin |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 58 | Qwen3.8-Max | independent |
| Artificial Analysis Intelligence Index | 56 | Qwen3.8-Flash-Next | independent |
| Artificial Analysis Intelligence Index | 52 | Qwen3.8-27B (xhigh) | independent |
| SWE-Bench Pro | 62.5 | Qwen3.8-Flash-Next | vendor-reported |
| LiveCodeBench v6 | 91.9 | Qwen3.8-Flash-Next | vendor-reported |
| CoWorkBench | 73.9 | Qwen3.8-Flash-Next | vendor-reported |
| OSWorld-Verified | 86.1 | Qwen3.8-Max | vendor-reported |
| SWE-Bench Pro | 67.7 | Qwen3.8-Max | vendor-reported |
| Terminal-Bench 2.1 | 86.6 | Qwen3.8-Max | vendor-reported |
| GPQA Diamond | 89.2 | Qwen3.8-27B | vendor-reported |
| OSWorld-Verified | 84.3 | Qwen3.8-27B | vendor-reported |
| SWE-Bench Pro | 61.7 | Qwen3.8-27B | vendor-reported |
| MMLU | 90.6% | Qwen3-235B VL | vendor-reported |
An independent index score for Qwen3.8-27B has been available since 17 August 2026 (52 points, xhigh mode). We still recommend using benchmark leaderboards for shortlisting and then measuring the finalists against your own representative tasks – especially because for this model the score is bought with very high token consumption, so cost per task diverges from headline token prices.
Hardware Requirements (Self-Hosted)
| Model | VRAM | Recommended GPU |
|---|---|---|
| Qwen3.8-2.4T-A95B | data-centre cluster | multi-node (H100/B200 class) |
| Qwen3.8-Flash-Next (FP8) | from approx. 180 GB for the weights | multi-GPU node (H100/H200 class); per press reports the n-gram embedding can also be served from system RAM |
| Qwen3.8-27B (quantised) | from approx. 24 GB | RTX 3090 / RTX 4090 / Apple Silicon with sufficient memory |
| Qwen3.5-397B-A17B | 80+ GB | H100/MI300X |
| Qwen3-235B-A22B | 48+ GB | A100/H100 |
| Qwen-Image-2.1 / -2.1-Turbo (7B, BF16) | from about 16 GB | RTX 4090 / L4 (non-commercial only) |
| Qwen-Image 20B | 48+ GB | A100/H100 |
| Qwen3 (smaller variants) | 16-24 GB | RTX 4090 |
Integration with CompanyGPT
Qwen models can be integrated into CompanyGPT as a self-hosted option – Qwen3.8-27B is a particularly good fit because it is Apache 2.0 licensed and already runs on a single 24 GB GPU. All processing stays in your infrastructure, with no third-country transfer.
Deprecations & Model Lifecycle
Alibaba Cloud Model Studio retires older models on a schedule. Under its current policy, Alibaba announces retirements at least three months ahead for stable models and at least 14 days ahead for preview models; from the start of retirement, new users no longer get access. On 10 October 2026, Alibaba shut down qwen3.6-max-preview, qwen3-max, qwen3-coder-plus, qwen3-coder-next, several qwen3-vl variants and the Qwen3 open-source models (incl. qwen3-32b, qwen3-235b-a22b, qwen3-next-80b-a3b), among others; Alibaba names qwen3.7-max, qwen3.6-flash and qwen3.7-plus as successors. A further retirement round has been announced for 19 January 2027.
The hyperscalers look similar: Google shut down all Qwen MaaS endpoints on 21 October 2026. On AWS Bedrock, Qwen3-Coder-30B-A3B-Instruct and Qwen3-Coder-480B-A35B-Instruct have been Legacy since 30 September 2026 (end of life 30 March 2027), while the other Qwen3 models remain active there.
Within the Qwen family, the preview builds Qwen3.5-Max-Preview and Qwen3.6-Max-Preview have been superseded by the GA models of the 3.7 and 3.8 lines. Qwen3.8-Flash-Next is explicitly not a GA model but an experimental preview; expect changes up to the Qwen4 generation. If you work via the API, pin model IDs to versions and actively monitor the sunset notices in Model Studio. Self-hosting avoids the issue entirely: once downloaded, Apache 2.0 weights remain usable regardless of the vendor’s roadmap – an often underrated argument for sovereign operation.
Our Recommendation
With Qwen 3.8, Alibaba draws a clear line between a freely usable working model, a licence-fenced flagship and an experimental architecture preview. For DACH enterprises, we recommend:
- Qwen3.8-27B: our top pick for self-hosting – Apache 2.0, multimodal, 262k context, runs on a 24 GB GPU, and now independently measured as well (52 points on the Artificial Analysis Intelligence Index). The best starting point for most enterprise scenarios; configure
reasoning_effortdeliberately. - Qwen3.6-35B-A3B: when maximum throughput per GPU hour matters – MoE with only 3B active parameters, also Apache 2.0.
- Qwen3.8-Flash-Next: for early evaluation of the Qwen4 architecture – 6B active parameters, multimodal, independently measured at 56 points; review the qwen-community-1.0 licence before commercial use.
- STACKIT AI Model Serving: Qwen3.8-27B as a managed, OpenAI-compatible API in region eu01 if you do not want to run it yourself.
- Qwen3.8-Flash via Model Studio (Frankfurt, scope “EU”): managed API with 1M context and inference within the EU; the provider remains Alibaba Cloud, so clarify the third-country access questions (see the GDPR section) first.
- Qwen3.8-Max via Model Studio (Frankfurt, scope “Global”): for tasks that need 1M context and high performance – inference may take place outside the EU, so assess the third-country transfer first.
- Qwen3.8-2.4T-A95B self-hosted: only with a reviewed licence position and data-centre hardware; the open weights are a text-only model.
- AWS Bedrock Frankfurt: for managed Qwen3 models such as Qwen3-32B or Qwen3-235B-A22B-2507 with EU data residency, where the 3.8 generation is not strictly required.
Which route fits you depends less on benchmark rankings than on your data-protection requirements and existing infrastructure. That is exactly the trade-off we help you work through – from model selection to production operation.