Qwen 2.5 and Qwen 3: Alibaba's open-weight family with maths and code strength

What is Qwen?

Qwen (from Chinese "Tongyi Qianwen", "1000-fold answer") is Alibaba Cloud's open-weight language model family. First releases in 2023, with Qwen 2.5 as the mature family since summer 2024 and Qwen 3 as the next generation since early 2026.

The Qwen family is broad. As of May 2026, Qwen 2.5 includes models in sizes 0.5B, 1.5B, 3B, 7B, 14B, 32B and 72B, each in base and instruct variants. Qwen 3, released in the first wave in March 2026, additionally offers MoE variants (Qwen3-30B-A3B with 3B active parameters, Qwen3-235B-A22B with 22B active) and improved reasoning via a "Thinking Mode" – similar to the DeepSeek-R1 or o3-mini approach, where the model runs an extensive reasoning step before the actual answer.

Licence situation May 2026: models up to and including Qwen2.5-72B are Apache 2.0. Models above 72B (such as Qwen2.5-72B-Plus, some Qwen 3 premium variants) sit under the Tongyi Qianwen License, a custom licence with commercial use up to 100 million monthly active users. For Swiss SMEs the 100M threshold is irrelevant – both licences are commercially usable in practice.

Availability: Hugging Face (Qwen/Qwen2.5-72B-Instruct, Qwen/Qwen3-30B-A3B-Instruct etc.), Alibaba Cloud DashScope API (with Chinese or Singapore hosting), Together AI, Fireworks AI, plus self-host via vLLM, TGI, Ollama, llama.cpp.

The Qwen family additionally includes specialised variants: Qwen2.5-Coder for programming (near top frontier models on SWE-Bench as of May 2026), Qwen2.5-Math for mathematics (top among open-weight on MATH and AIME), QwenVL for vision-language and Qwen-Audio for speech processing. This specialisation is a clear differentiator in May 2026.

Why Qwen matters for Swiss data

Qwen has four concrete arguments and two important caveats for Swiss setups as of May 2026.

First: maths and code as class-best. Qwen2.5-Math and Qwen2.5-Coder clearly beat all other open-weight families on the respective benchmarks. For a fiduciary firm with complex tax calculation pipelines (special VAT cases, international transfer pricing, pension fund mathematics), Qwen2.5-Math is a productive help. For an internal tool development team having scripts and microservices generated, Qwen2.5-Coder is on the level of Claude Sonnet in the code area.

Second: multilingual strength and Qwen3 Thinking Mode. Qwen is trained competently in around 30 languages, with particular focus on Mandarin (logical for Alibaba), English, German, French, Spanish and Japanese. On German, Qwen is productively usable in May 2026, though not quite at Mistral level. Qwen3 Thinking Mode delivers results on hard logic tasks close to frontier models – interesting for complex fiduciary or legal reasoning cases.

Third: Apache 2.0 for relevant sizes. Qwen 2.5 up to 72B as Apache 2.0 is the cleanest licence configuration. Self-host without commercial restrictions, fine-tuning allowed, modification allowed. Attractive for SME compliance setups.

Fourth: specialised models cover specific use cases optimally. Whoever needs receipt-photo processing with a vision-language model has a top option in QwenVL. Whoever needs internal code assistance has a premium variant with Qwen2.5-Coder. This specialisation saves the search for individual best-of-breed models.

Caveat one: PRC origin. Alibaba is a Chinese company. With direct API use through DashScope, requests go either to the mainland China data centre or the Singapore centre. For Swiss clients under professional secrecy per Art. 321 SCC this is excluded – and for GDPR-compliant setups too, third-country transfer entails additional TIA duties. Self-host via Hugging Face solves the problem: the weights are open-weight and run in own rack without any request reaching Alibaba.

Caveat two: political risk and sanctions situation. The US-EU-PRC relationship is tense in May 2026. Future sanctions restricting Hugging Face downloads of PRC models cannot be ruled out. Whoever builds on Qwen should have a backup strategy (secure model weights locally, keep Apertus or Mistral as plan-B models ready).

Qwen in practice

Architecture. Qwen 2.5 is a dense transformer-decoder model with grouped-query attention, rotary position embeddings and SwiGLU activation – very similar to the Llama family. Context window on main models 128k tokens. Qwen 3 introduces MoE variants (Qwen3-30B-A3B: 30B total, 3B active) and the Thinking Mode, activated via the /think prompt tag, which prompts the model to run an extensive internal reasoning step.

Setup example with Ollama. Qwen 2.5 72B on two H100 or one H100 with quantisation:

``` ollama pull qwen2.5:72b-instruct-q4_K_M ollama run qwen2.5:72b-instruct-q4_K_M "Calculate Swiss VAT on EUR 12500 at 8.1 percent." ```

Performance: on two H100 80GB in 4-bit AWQ around 30-50 tokens/s, on one H100 with GGUF Q4_K_M around 15-25 tokens/s.

Setup example with vLLM. Qwen3-30B-A3B (MoE) on one H100:

``` docker run --gpus all -p 8000:8000 \ vllm/vllm-openai:v0.6.3 \ --model Qwen/Qwen3-30B-A3B-Instruct \ --max-model-len 32768 \ --gpu-memory-utilization 0.92 \ --enable-prefix-caching ```

MoE architecture means: 30B total parameters, but only 3B active per token. Result: low inference cost on a single H100, comparable to a dense 3B model – but quality more on the 14B level.

Thinking Mode Qwen 3. Activation via system prompt:

``` System: /think User: Solve this task step by step: A company has 12 employees, of whom 3 are part-time. Per full-time position CHF 500 pension contribution per month can be claimed. What is the annual sum? ```

The model first generates a "thinking block" with extensive reasoning, then the final answer. The thinking block is visible in the output and can be logged for audit purposes – important for EU AI Act Art. 15 logging duties.

Code workflow with Qwen2.5-Coder. Qwen2.5-Coder exists in 7B, 14B and 32B. On an RTX 4090, Qwen2.5-Coder-32B runs in 4-bit AWQ quantisation at 30-50 tokens/s – productive for internal code assistance. Integration via Continue.dev or VS Code extension with an OpenAI API endpoint.

RAG setup with Qwen embeddings. Qwen has its own embedding models (Qwen3-Embedding-8B leading on MTEB benchmark as of May 2026). In a Swiss RAG pipeline, Qwen3-Embedding-8B can be loaded into LocalAI or TEI, with Qdrant as the vector DB and Apertus 70B or Mistral Large 2 as the generator model.

Hosting recommendation. Apache 2.0 Qwen models run best self-host: Hugging Face download once, then fully offline-capable. For EU/CH compliance, this variant is the clean choice. DashScope API via Singapore can be an option for non-sensitive workloads, but a GDPR TIA and FADP third-country review are mandatory.

Qwen to production in 5 steps

1. 01Compliance check: assess PRC origin in the context of client sensitivity and compliance policy. Self-host is the clean variant; API use via DashScope requires TIA and FADP third-country review.
2. 02Model choice: Qwen2.5-72B as general workhorse (Apache 2.0), Qwen2.5-Math for maths, Qwen2.5-Coder-32B for code, Qwen3-30B-A3B for efficient MoE inference, Qwen3 Thinking variant for complex reasoning.
3. 03Hardware check: Qwen2.5-72B in 4-bit AWQ needs ca. 45 GB VRAM (one H100 or two RTX 4090). MoE variants are memory-efficient – Qwen3-30B-A3B fits an RTX 4090.
4. 04Self-host via vLLM or Ollama, OpenAI-compatible endpoint, LiteLLM in front with logical model names (qwen-math-local, qwen-coder-local, qwen3-thinking-local).
5. 05Use-case benchmark against Apertus 70B, Mistral Large 2 and the current top Claude model. Derive routing rules from it: maths-intensive queries to Qwen-Math, code generation to Qwen-Coder, sensitive CH language to Apertus, general reasoning to Mistral or Claude.

When to use Qwen

Qwen is the right choice when (a) maths or code-specialised workloads are central, (b) Apache 2.0 licence is central, or (c) Qwen3 Thinking Mode is needed for complex reasoning.

Concrete cases: fiduciary with complex tax calculation pipelines – Qwen2.5-Math as self-host for maths-intensive workloads. Software consulting boutique with internal code generation – Qwen2.5-Coder-32B on an RTX 4090. SME with a RAG setup wanting to keep embeddings local – Qwen3-Embedding-8B in LocalAI or TEI.

For Swiss setups that exclusively self-host and prioritise an Apache 2.0 licence, Qwen 2.5 up to 72B is a direct competitor to Apertus 70B. Apertus is ahead on CH-specific language; Qwen is ahead on maths and code. A multi-provider strategy with both is sensible.

When not to use

For setups with highly sensitive client data (professional secrecy, strict FINMA mandates), the DashScope API variant is excluded – even Singapore hosting remains third-country. Self-host stays open, but the supra-political argument (PRC origin) must be addressed in internal discussion. Whoever does not want this goes to Apertus, Mistral or Llama 4 as alternatives.

For Romansh or Schwizerdütsch workloads, Qwen is not trained. Apertus remains the right choice here.

For top frontier reasoning at the peak (math olympiad level beyond Qwen2.5-Math, complex legal four-step argument), the current top Claude model or the current top GPT model are still ahead. Qwen3 with Thinking Mode is close but not quite at frontier level.

For setups where a US or EU provider commitment is desired for compliance reasons (e.g. a Swiss bank that only accepts Western providers), Qwen is the wrong choice irrespective of technical quality.

Trade-offs

FAQ

Related topics

Sources

1. Qwen – official model collection on Hugging Face · 2026-05
2. Qwen 3 – release notes and Thinking Mode introduction · 2026-03
3. Qwen2.5-Math – paper and benchmarks · 2026-04
4. Alibaba Cloud DashScope – Qwen API documentation · 2026-05