Qwen (Alibaba)

From Systems Analysis Wiki
Jump to navigation Jump to search

Qwen (Chinese: 通义千问; pinyin: Tōngyì Qiānwèn) is a family of artificial-intelligence foundation models developed by the Qwen team at Alibaba Cloud, part of Alibaba Group. The family includes text-oriented large language models (LLMs), coding and mathematical models, vision-language and omni-modal systems, image-generation models, speech-recognition and speech-synthesis models, embedding and reranking models, safety classifiers, language world models, and models intended for agents and robotics.[1]

Alibaba announced a beta of the Tongyi Qianwen service in April 2023. The first downloadable weights, Qwen-7B and Qwen-7B-Chat, were released in August 2023, while the hosted service was opened to the Chinese public in September 2023.[2][3]

Qwen follows a mixed distribution strategy. Alibaba releases weights for many models, frequently under the Apache License 2.0, while other checkpoints use Qwen-specific licenses or are available only as hosted Alibaba Cloud services. The complete training corpora and reproducible end-to-end training pipelines are not published for the frontier models. Consequently, the family as a whole should not be described as uniformly "open source"; open-weight is the more precise term for checkpoints whose weights can be downloaded.[4]

As of July 11, 2026, Alibaba Cloud Model Studio recommends Qwen3.7-Max, Qwen3.7-Plus, and Qwen3.6-Flash for general text generation. Qwen3.7-Max is the proprietary text-focused flagship, while Qwen3.7-Plus adds image and video understanding. Both support contexts of up to one million tokens in the documented hosted service. The newest openly released general-purpose checkpoints are Qwen3.6-27B and Qwen3.6-35B-A3B; newer specialized open releases include Qwen-AgentWorld and the Qwen-Robot suite.[5][6][7]

History and development

The Qwen family originated in Alibaba Cloud's broader generative-AI program. Tongyi Qianwen was initially introduced as a proprietary cloud service. In August 2023, Alibaba released Qwen-7B and Qwen-7B-Chat; Qwen-14B, Qwen-1.8B, and Qwen-72B followed later that year. The code repository used Apache 2.0, but the model weights were governed by Tongyi Qianwen licenses rather than Apache 2.0. Commercial use of the 7B, 14B, and 72B weights required an application, while Qwen-1.8B used a research license and required separate contact for commercial use.[8][9]

Qwen2, released in June 2024, introduced a broader set of dense models and the Qwen2-57B-A14B Mixture-of-Experts model. Its instruction-tuned checkpoints supported context windows ranging from 32K to 128K, depending on model size. Qwen2-72B retained the Qianwen License, while the 0.5B, 1.5B, 7B, and 57B-A14B models adopted Apache 2.0.[10][11]

Qwen2.5, released in September 2024, expanded the reported pre-training corpus to as much as 18 trillion tokens and improved coding, mathematics, structured-data handling, long-text generation, and instruction following. Its general-purpose line included 0.5B, 1.5B, 3B, 7B, 14B, 32B, and 72B models. The 3B and 72B checkpoints used the Qwen Research License and Qwen License, respectively; the other general-purpose sizes used Apache 2.0.[12][13]

The separate QwQ line explored reinforcement-learning-based mathematical and logical reasoning. QwQ-32B-Preview was released in November 2024, followed by the final QwQ-32B in March 2025. The final model incorporated agent-related capabilities and was released under Apache 2.0.[14]

Qwen3 was released on April 29, 2025 with six dense and two MoE models. The original release combined thinking and non-thinking operation within the same post-trained checkpoints, used approximately 36 trillion pre-training tokens, and supported 119 languages and dialects. All eight original Qwen3 text checkpoints used Apache 2.0.[15][16][17]

In July and August 2025, Alibaba released the Qwen3-2507 updates. Unlike the initial hybrid checkpoints, these were distributed as separate Instruct and Thinking variants. The 235B-A22B and 30B-A3B updates supported native 256K contexts and optional extension to approximately one million tokens. Specialized 2025 releases included Qwen3-Coder, Qwen3-VL, Qwen3-Omni, Qwen3-Embedding, Qwen3-Reranker, and Qwen3Guard.[17]

In September 2025, Qwen3-Next-80B-A3B introduced an ultra-sparse MoE design with approximately 80 billion total and 3 billion active parameters. It combined Gated DeltaNet linear-attention blocks with gated full-attention blocks and became the architectural basis for several later Qwen models.[18][19]

Alibaba introduced the proprietary Qwen3-Max in September 2025. The company described it as an MoE model with more than one trillion total parameters and approximately 36 trillion pre-training tokens. Qwen3-Max-Instruct was followed by Qwen3-Max-Thinking in January 2026. Neither version was released as a downloadable open-weight checkpoint.[20][21]

In February 2026, Qwen3.5 introduced a native vision-language mainline architecture based on early text-vision fusion. The family combines Gated DeltaNet linear attention with gated full attention and contains both dense and sparse-MoE checkpoints. Open models ranged from Qwen3.5-0.8B to Qwen3.5-397B-A17B. Alibaba reported support for 201 languages and dialects and expanded the vocabulary to approximately 250,000 tokens.[22][7]

Qwen3.6-Plus, a proprietary hosted model, was released on April 2, 2026. Open weights followed for Qwen3.6-35B-A3B on April 16 and Qwen3.6-27B on April 22. Both open models are native vision-language systems, support thinking and non-thinking operation, and have native contexts of 262,144 tokens. The model cards describe optional extension to 1,010,000 tokens through long-context configuration.[23][24][25]

Alibaba introduced the proprietary Qwen3.7-Max on May 21, 2026 and the multimodal Qwen3.7-Plus on June 3. Max focuses on long-horizon coding, productivity, and tool-using agents. Plus accepts text, images, and video, and adds graphical-user-interface operation, visual grounding, and coding from visual references. The documented hosted versions support one-million-token contexts.[26][27][6]

June 2026 also brought two research-oriented expansions:

  • Qwen-Robot Suite, containing Qwen-RobotManip for vision-language-action control, Qwen-RobotNav for vision-language navigation, and Qwen-RobotWorld for embodied video-world modelling.[28]
  • Qwen-AgentWorld, a language-world-model project for simulating agent environments. The accompanying paper evaluates 35B-A3B and 397B-A17B models, but the publicly released checkpoint is Qwen-AgentWorld-35B-A3B, together with AgentWorldBench.[29][30]
Major Qwen Model Releases
Release Date Model Parameters (B) Key Features License
April 2023 Tongyi Qianwen service Not disclosed Initial beta of Alibaba Cloud's hosted conversational and enterprise model service.[1] Proprietary cloud service
August–November 2023 Qwen 1 1.8, 7, 14, and 72 Text models trained on approximately 2.2–3.0 trillion tokens. Context limits differed by checkpoint and update: 8K for Qwen-14B and up to 32K for Qwen-1.8B, Qwen-7B, and Qwen-72B.[8][9] Tongyi Qianwen model-specific licenses
June 2024 Qwen2 Dense: 0.5, 1.5, 7, and 72;
MoE: 57B-A14B 		GQA throughout the family; 27 explicitly added languages beyond Chinese and English. Instruction-model contexts were 32K, 32K, 128K, 64K, and 128K, respectively.[10] Apache 2.0 except Qwen2-72B, which uses the Qianwen License
September 2024 Qwen2.5 0.5, 1.5, 3, 7, 14, 32, and 72 Up to 18 trillion pre-training tokens; stronger coding, mathematics, structured-data analysis, JSON generation, and long-text generation. Contexts of 32K or 128K depending on size.[12][13] Apache 2.0 for 0.5B, 1.5B, 7B, 14B, and 32B; Qwen Research License for 3B; Qwen License for 72B
November 2024 / March 2025 QwQ-32B Preview / QwQ-32B 32 Reasoning-oriented models developed with reinforcement learning. The final release added stronger agent and tool-use capabilities.[14] Apache 2.0
January–March 2025 Qwen2.5-VL / Qwen2.5-Omni VL: 3, 7, 32, and 72;
Omni: 7 		Qwen2.5-VL processes text, images, and video. Qwen2.5-Omni accepts text, image, audio, and video and streams text or speech through a Thinker–Talker architecture.[31][32] Checkpoint-dependent; Qwen2.5-Omni-7B uses Apache 2.0
April 2025 Qwen3 Dense: 0.6, 1.7, 4, 8, 14, and 32;
MoE: 30B-A3B and 235B-A22B 		Approximately 36 trillion pre-training tokens; 119 languages and dialects; unified thinking/non-thinking operation; contexts up to 128K in the initial release.[15][16] Apache 2.0
July–August 2025 Qwen3-2507 4, 30B-A3B, and 235B-A22B Separate Instruct and Thinking checkpoints; native 256K long-context support for the larger variants and optional extension to approximately one million tokens.[17] Apache 2.0
July–September 2025 Qwen3-Coder, Qwen3-VL, Qwen3-Omni Model-dependent Coding, vision-language, and end-to-end omni-modal families. Qwen3-Omni processes text, image, audio, and video and can stream text and natural speech.[33][34][35] Primarily Apache 2.0 for published weights; verify the individual model card
September 2025 Qwen3-Next / Qwen3-Max Next: 80B-A3B;
Max: more than 1T total 		Qwen3-Next introduced an ultra-sparse MoE and hybrid Gated DeltaNet/full-attention architecture. Qwen3-Max was a proprietary hosted frontier model.[18][20] Next: Apache 2.0;
Max: proprietary cloud service
February–March 2026 Qwen3.5 Dense: 0.8, 2, 4, 9, and 27;
MoE: 35B-A3B, 122B-A10B, and 397B-A17B 		Native vision-language models; hybrid Gated DeltaNet and gated attention; dense and sparse-MoE variants; 201 languages and dialects; approximately 250K-token vocabulary.[22][7] Apache 2.0 for open checkpoints; proprietary Qwen3.5-Plus service
April 2026 Qwen3.6 Dense: 27;
MoE: 35B-A3B 		Native vision-language open models emphasizing agentic coding and reasoning. Native 262,144-token context, extendable through configuration to approximately 1.01 million tokens.[36][37] Apache 2.0
May–June 2026 Qwen3.7-Max / Qwen3.7-Plus Not disclosed Proprietary agent models. Max is text-focused; Plus accepts text, images, and video. Both provide one-million-token contexts in the documented hosted service.[26][27][6] Proprietary cloud service
June 2026 Qwen-AgentWorld-35B-A3B 35 total / 3 active Language world model with a 256K context for simulating MCP, search, terminal, software-engineering, Android, web, and operating-system environments; trained using more than 10 million interaction trajectories.[29][30] Apache 2.0
June 2026 Qwen-Robot Suite Model-dependent Qwen-RobotManip for robotic manipulation, Qwen-RobotNav for navigation, and Qwen-RobotWorld for embodied world modelling.[28] Verify the individual model or repository

Architecture and technical features

Qwen does not use one unchanged architecture across all generations. Qwen 1 through the original Qwen3 text models primarily use autoregressive causal decoder-only Transformer architectures. Vision-language models add visual encoders and modality adapters. Qwen3-Next, Qwen3.5, and Qwen3.6 use hybrid language backbones combining linear-attention and full-attention blocks. Omni-modal, speech, image-generation, embedding, reranking, world-model, and robotics systems use additional specialized components.[11][15][22]

Key Architectural Components

  • Core Transformer Elements: Earlier Qwen text families use pre-normalization with RMSNorm, gated SwiGLU feed-forward layers, causal next-token prediction, and residual connections. Qwen3 added QK normalization and revised several attention, routing, and normalization details.[11][15]
  • Positional Encoding: Qwen text models use Rotary Positional Embeddings (RoPE). Different generations have supplemented RoPE with methods including NTK-aware interpolation, LogN scaling, Dual Chunk Attention, and YaRN (Yet Another RoPE Extension). No single technique is responsible for every reported context limit.[8][11][38]
  • Grouped Query Attention: Qwen2 adopted Grouped Query Attention (GQA) across all five principal model sizes. GQA uses fewer key/value heads than query heads, reducing the KV-cache size and memory bandwidth required during autoregressive generation.[10]
  • Hybrid Linear and Full Attention: Qwen3-Next, Qwen3.5, and Qwen3.6 interleave Gated DeltaNet linear-attention blocks with gated full-attention blocks. For example, Qwen3.6 uses a repeated pattern of three Gated DeltaNet blocks followed by one gated-attention block.[22][36][37]
  • Multi-Token Prediction: Later models are trained to predict multiple future tokens during training. The auxiliary objective is intended to improve representation learning and can also support speculative or accelerated decoding implementations.[22][36]
  • Implementation Kernels: Inference and training frameworks may use kernels such as FlashAttention. FlashAttention is an implementation technique for exact attention rather than a unique learned architectural component of Qwen.

Dense Models and Mixture of Experts (MoE)

The Qwen family contains both dense and sparse-MoE models:

  • Dense Models: Nearly all language-model parameters participate in processing each token. Examples include Qwen2-72B, Qwen2.5-32B, Qwen3-32B, Qwen3.5-27B, and Qwen3.6-27B.
  • Mixture-of-Experts Models: Feed-forward computation is divided among multiple experts, and a learned router activates only a subset for each token. Total parameters and active parameters should therefore be reported separately.
    • Qwen2-57B-A14B contains approximately 57 billion total parameters and activates approximately 14 billion.[11]
    • Qwen3-30B-A3B and Qwen3-235B-A22B contain approximately 30/235 billion total parameters and activate approximately 3/22 billion, respectively. The original Qwen3 MoE design uses 128 routed experts and selects eight for each token.[16]
    • Qwen3-Next-80B-A3B contains approximately 80 billion total parameters and activates approximately 3 billion.[18]
    • Qwen3.5-35B-A3B, Qwen3.5-122B-A10B, and Qwen3.5-397B-A17B activate approximately 3, 10, and 17 billion parameters, respectively.[22][7]
    • Qwen3.6-35B-A3B contains approximately 35 billion total parameters and activates approximately 3 billion. Its model card specifies 256 routed experts, with eight routed experts and one shared expert participating for each token.[37]

Sparse activation reduces arithmetic relative to a dense model with the same total parameter count. It does not reduce storage requirements to the active-parameter count: all expert weights generally need to remain stored or accessible across accelerators. Real serving efficiency also depends on memory bandwidth, expert routing, batching, parallelism, precision, sequence length, and inference software.

Innovations for Long Context

Context limits vary by model, checkpoint, runtime configuration, and hosted endpoint.

  • Qwen 1: Qwen-14B supported approximately 8K tokens. Updated Qwen-1.8B and Qwen-7B configurations, and Qwen-72B, supported up to 32K.[8]
  • Qwen2: Instruction-model contexts were 32K for 0.5B and 1.5B, 128K for 7B, 64K for 57B-A14B, and 128K for 72B. The 7B and 72B long-context results used YaRN-assisted extension.[10]
  • Qwen2.5: The 0.5B, 1.5B, and 3B general models used 32K contexts; the 7B, 14B, 32B, and 72B models supported 128K. Separate Qwen2.5-1M variants of the 7B and 14B instruction models were released in January 2025.[12][39]
  • Original Qwen3: The 0.6B, 1.7B, and 4B models supported 32K, while 8B, 14B, 32B, 30B-A3B, and 235B-A22B supported 128K.[16]
  • Qwen3-2507 and Qwen3-Next: The larger Qwen3-2507 checkpoints and Qwen3-Next support native contexts of approximately 256K. Selected checkpoints can be configured for approximately one million tokens.[17][18]
  • Qwen3.5 and Qwen3.6 open models: Model cards specify a native context of 262,144 tokens and optional extension to 1,010,000 tokens. The extension requires appropriate YaRN or runtime configuration and is not necessarily enabled by default in every framework.[36][37]
  • Hosted models: Qwen3.5-Plus, Qwen3.6-Plus, Qwen3.6-Flash, Qwen3.7-Max, and Qwen3.7-Plus provide one-million-token contexts in their documented Alibaba Cloud configurations. Qwen3.7-Plus supports up to 64K output tokens.[22][23][6]

A nominal context window states the maximum accepted sequence length. It does not guarantee uniform retrieval, comprehension, reasoning, or factual consistency across every prompt of that length.

"Thinking Mode" in Qwen 3 and Later Models

The initial Qwen3 post-trained models introduced a hybrid reasoning design:

  • Thinking mode generates additional internal reasoning tokens before the final response.
  • Non-thinking mode produces a more direct response with lower latency.
  • A thinking budget or token limit can constrain inference-time reasoning.

Mode selection is controlled through the chat template, API parameter, or client. Open Qwen3 deployments can use /think and /no_think where supported. Alibaba Cloud's later hosted models generally use the enable_thinking parameter. Qwen3.5, Qwen3.6, Qwen3.7-Max, and Qwen3.7-Plus support hybrid thinking in documented endpoints.[16][40]

The behaviour is not uniform across all specialized models. Qwen3-2507 was distributed as separate Instruct and Thinking checkpoints, and Qwen3-Coder-Next is a non-thinking model. Some applications expose reasoning text in <think>...</think> tags or a separate API field, while others suppress it. It is therefore inaccurate to state that every Qwen response necessarily displays the model's complete reasoning process.[17][33]

Multilingual Tokenizer

There is no single unchanged tokenizer across the complete family.

  • Qwen 1 used a tokenizer implemented with the tiktoken framework, together with Qwen-specific vocabulary and special tokens.[8]
  • Qwen2 uses byte-level BPE with 151,643 regular tokens and three control tokens. It should not be described simply as a copy of OpenAI's cl100k vocabulary.[11]
  • Qwen3 retained a vocabulary of approximately 152,000 tokens while expanding model support to 119 languages and dialects.[15]
  • Qwen3.5 and Qwen3.6 increased the padded token vocabulary to 248,320 entries, commonly described as approximately 250,000 tokens. Alibaba reported improved encoding efficiency for many languages.[22][36]

A larger vocabulary can reduce the number of tokens required for some scripts and programming languages, but tokenization efficiency still varies by language, writing system, and text domain.

Multimodal Capabilities

The Qwen family contains several distinct multimodal branches and architectures:

  • Qwen-VL, Qwen2-VL, Qwen2.5-VL, and Qwen3-VL: Vision-language models for image and video understanding, visual question answering, optical character recognition, document analysis, spatial grounding, and visual-agent tasks. Qwen2.5-VL and Qwen3-VL support dynamic-resolution inputs. Qwen3-VL includes dense 2B, 4B, 8B, and 32B variants and MoE 30B-A3B and 235B-A22B variants, with native interleaved contexts of up to 256K.[31][41]
  • Qwen-Audio and Qwen2-Audio: Audio-language models for understanding speech, music, and environmental sounds. Their principal output is text; they should not be conflated with end-to-end speech-generation systems.[42]
  • Qwen2.5-Omni and Qwen3-Omni: End-to-end systems accepting combinations of text, image, audio, and video and producing text or natural speech. Qwen2.5-Omni introduced the Thinker–Talker organization, while Qwen3-Omni expanded the open omni-modal line with a 30B-A3B architecture.[32][35]
  • Qwen3.5-Omni: A large omni-modal model using a hybrid-attention MoE Thinker–Talker architecture. Its technical report describes a 256K model context and support for text, images, audio, video, text generation, and speech generation. Hosted service limits can differ: Alibaba Cloud's current documentation lists a 64K context for qwen3.5-omni-plus.[43][6]
  • Native multimodal mainline models: Qwen3.5 and Qwen3.6 integrate visual and textual processing into the main model family. Their open model cards classify them as causal language models with vision encoders and support text, image, and video workflows.[36][37]
  • Qwen3.7-Plus: A proprietary multimodal agent model accepting text, images, and video and producing text. The hosted service supports visual grounding, graphical-interface operation, function calling, built-in tools, up to two-hour video inputs, and a one-million-token context.[27][6]
  • Image generation: Qwen-Image and related editing checkpoints generate or modify images from text and visual references. Image-generation models are architecturally and functionally distinct from the vision-language models used primarily for understanding.[44]
  • Speech-specialized models: Qwen3-ASR provides multilingual speech recognition, language identification, and timestamp prediction; Qwen3-TTS provides speech synthesis, voice design, and voice-cloning functions.[45][46]
  • Retrieval and safety models: Qwen3-Embedding and Qwen3-Reranker provide 0.6B, 4B, and 8B checkpoints for multilingual retrieval and ranking. Qwen3Guard provides safety-classification models. These models do not perform the same function as a general conversational LLM.[47]

Training Data and Scale

Qwen technical reports describe training mixtures containing web text, books, code, mathematics, scientific material, multilingual text, document-derived content, visual data, audio, video, and synthetic examples. The exact source list, complete raw corpus, document-level provenance, filtering pipeline, and reproducible end-to-end training procedure have not been published for the frontier models.

  • Qwen 1: Approximately 2.2 trillion tokens for Qwen-1.8B, 2.4 trillion for the updated Qwen-7B, and 3.0 trillion for Qwen-14B and Qwen-72B.[8]
  • Qwen2: The principal corpus was expanded from approximately 3 trillion to approximately 7 trillion tokens. Training amounts and continued-training stages differed by checkpoint.[11]
  • Qwen2.5: Up to approximately 18 trillion tokens, followed by supervised fine-tuning on more than one million examples and multi-stage preference or reinforcement-learning procedures.[13]
  • Qwen3: Approximately 36 trillion tokens covering 119 languages and dialects. The team reported using Qwen2.5-VL to extract text from PDF-like documents and Qwen2.5-Coder and Qwen2.5-Math to generate or improve specialized synthetic data.[16]
  • Qwen3-Max: Alibaba reported more than one trillion parameters and approximately 36 trillion pre-training tokens.[20]
  • Qwen3.5: Alibaba described substantially expanded visual-text training, stricter data filtering, 201-language coverage, a native FP8 training pipeline, and reinforcement learning across large numbers of agent environments. It did not publish one complete corpus-size figure or the underlying training dataset.[22]
  • Qwen3.6: The open model cards describe pre-training and post-training but do not disclose a separate complete corpus size. The models build on the Qwen3.5 architecture and training program.[36][37]
  • Qwen-AgentWorld: The environment-modelling pipeline used more than 10 million real-world interaction trajectories across seven agent domains. This is specialized mid-training and post-training data rather than the full corpus used to create the underlying Qwen3.5 base model.[29][30]

Reported token counts are not directly comparable without information about tokenization, deduplication, repeated data, modality conversion, synthetic-data generation, filtering, and curriculum design. A larger token count does not by itself establish higher data quality or better production performance.

Licensing and Availability

Licensing varies by generation, size, and checkpoint:

  • Qwen 1: Repository code uses Apache 2.0. Qwen-7B, Qwen-14B, and Qwen-72B weights use the Tongyi Qianwen License and historically required an application for commercial use. Qwen-1.8B uses a research license.[8]
  • Qwen2: Qwen2-0.5B, 1.5B, 7B, and 57B-A14B use Apache 2.0. Qwen2-72B uses the Qianwen License.[10]
  • Qwen2.5: The 0.5B, 1.5B, 7B, 14B, and 32B general models use Apache 2.0. Qwen2.5-3B uses the Qwen Research License, and Qwen2.5-72B uses the Qwen License. Vision-language and specialized checkpoints must be checked individually.[12]
  • Qwen3: The original eight text checkpoints use Apache 2.0. Qwen3-2507 and most published Qwen3-Coder, Qwen3-VL, Qwen3-Omni, Qwen3-Embedding, and Qwen3-Reranker weights also use Apache 2.0, although the exact model card remains authoritative.[16][17]
  • Qwen3-Next, Qwen3.5, and Qwen3.6: Published open-weight checkpoints use Apache 2.0.[7]
  • Qwen-AgentWorld: The released Qwen-AgentWorld-35B-A3B weights and AgentWorldBench use Apache 2.0.[29]
  • Proprietary services: Qwen-Plus, Qwen-Max, Qwen3-Max, Qwen3.5-Plus, Qwen3.6-Plus, Qwen3.6-Flash, Qwen3.7-Max, and Qwen3.7-Plus are hosted products rather than publicly downloadable open-weight checkpoints.[5][20][22][23][26][27]

Weights and code for open releases are distributed through Hugging Face, ModelScope, and GitHub. Hosted models are accessed through Alibaba Cloud Model Studio and related regional services. Availability, aliases, quotas, data-processing terms, context limits, and prices can differ by deployment region and can change over time.

Apache 2.0 permits use, modification, redistribution, and commercial deployment subject to its conditions. It does not provide the undisclosed training data or a reproducible training pipeline. Under definitions requiring those components, an Apache-licensed weight release can still be more precisely described as open-weight rather than fully open-source AI.[4]

Comparison with Competitors and Performance

Qwen benchmark comparisons are generally developer-reported. Results depend on the exact checkpoint, prompt template, evaluation harness, sampling settings, quantization, tool access, context-management strategy, reasoning budget, and scoring method.

  • vs. Llama (Meta): In the Qwen2 technical report, Qwen2-72B scored above Llama 3 70B on the listed MMLU, MMLU-Pro, HumanEval, and GSM8K evaluations. These were base-model results under Qwen's evaluation settings.[11]
  • vs. Mistral AI: In the same report, Qwen2-72B scored above Mixtral 8x22B on the four reproduced base-model benchmarks. This does not establish superiority across every task, language, context length, or deployment constraint.[11]
  • vs. proprietary models: Alibaba has reported competitive or leading results for later Qwen models on selected reasoning, coding, multimodal, and agent evaluations. Comparisons with changing commercial endpoints are particularly sensitive to date, tool access, inference-time compute, and the surrounding agent scaffold.[15][22][26][27]

Benchmark Results

Base-model results reproduced from the Qwen2 technical report (2024)[11]
Model MMLU (5-shot) MMLU-Pro (5-shot) HumanEval (0-shot) GSM8K (5-shot)
Qwen2-72B 84.2 55.6 64.6 89.5
Llama 3 70B 79.5 52.8 48.2 83.0
Mixtral 8x22B 77.8 49.5 46.3 83.7

The Qwen2 report separately listed Qwen2-72B-Instruct at 9.12 on MT-Bench, 48.1 on Arena-Hard, and 35.7 on LiveCodeBench. Those instruction-model evaluations used protocols different from the base-model table and are therefore not combined with it.[11]

Note: Benchmark values can change because of revised datasets, decontamination, prompt changes, inference settings, model updates, tool use, and agent scaffolding. Results reported by a model developer should be distinguished from independent testing and from performance in a specific production application.

Ecosystem and Application

The Qwen family is integrated into Alibaba's cloud, developer, application, and research ecosystem:

  • Qwen Studio: The public web and application interface formerly known as Qwen Chat. It provides conversational, document, image, video, generation, and agent-oriented functionality.[7]
  • Alibaba Cloud Model Studio: Provides hosted Qwen APIs and third-party models. The Qwen3.6 documentation describes compatibility with OpenAI-style and Anthropic-style API specifications. As of July 2, 2026, the recommended text models are qwen3.7-max, qwen3.7-plus, and qwen3.6-flash.[5][7]
  • PAI-EAS: Alibaba Cloud infrastructure for deploying, adapting, and serving models in managed or enterprise environments.
  • Open-weight Community: Checkpoints, model cards, examples, and quantizations are distributed through Hugging Face, ModelScope, and GitHub. Qwen models are supported to varying degrees by Transformers, vLLM, SGLang, llama.cpp, MLX, KTransformers, and other runtimes.[7]
  • Qwen-Agent: An open-source agent framework for function calling, retrieval, planning, memory, code interpretation, Model Context Protocol integration, and tool orchestration.[48]
  • Qwen Code: An open-source terminal-based coding agent optimized for Qwen models and repository-level software-engineering workflows.[49]
  • Qwen3-Coder-Next: An 80-billion-total, 3-billion-active open-weight coding model based on Qwen3-Next. It supports a native 256K context, optional extension to approximately one million tokens, and 358 programming languages. It was trained using executable tasks, environment interaction, and reinforcement learning.[50]
  • Qwen-AgentWorld: Provides an open language world model and benchmark for simulating MCP, search, terminal, software-engineering, Android, web, and operating-system environments. The project can be used to train or evaluate agents without requiring every rollout to interact with a live external system.[29]
  • Qwen-Robot Suite: Extends the Qwen ecosystem to physical AI. Qwen-RobotManip maps vision and language to robotic actions; Qwen-RobotNav focuses on embodied navigation; Qwen-RobotWorld models possible future physical states from video and action context.[28]
  • Applications: Qwen models are used for conversational assistance, programming, document processing, search, translation, data analysis, content generation, image and video understanding, speech interfaces, enterprise automation, software agents, simulated environments, and robotics research.

Model Context Protocol compatibility is generally supplied by Qwen-Agent, Qwen Code, or another surrounding agent framework. It is not an inherent network capability possessed by model weights in isolation.

Suitability depends on the selected checkpoint, license, deployment region, latency and hardware constraints, data-governance requirements, and application-specific evaluation. Local deployment can increase control over data location and system configuration but does not automatically establish accuracy, security, regulatory compliance, or safety.

Literature

  • Bai, Jinze; et al. (2023). Qwen Technical Report. arXiv:2309.16609.
  • Yang, An; et al. (2024). Qwen2 Technical Report. arXiv:2407.10671.
  • Chu, Yunfei; et al. (2024). Qwen2-Audio Technical Report. arXiv:2407.10759.
  • Qwen Team (2025). Qwen2.5 Technical Report. arXiv:2412.15115.
  • Bai, Jinze; et al. (2025). Qwen2.5-VL: A Versatile Vision-Language Model for Real-World Agent Tasks. arXiv:2502.13923.
  • Wang, Wen; et al. (2025). Qwen2.5-Omni: A Streaming End-to-End Multimodal Model. arXiv:2503.20215.
  • Yang, An; et al. (2025). Qwen3 Technical Report. arXiv:2505.09388.
  • Qwen Team (2025). Qwen3-VL Technical Report. arXiv:2511.21631.
  • Qwen Team (2026). Qwen3-Coder-Next Technical Report. arXiv:2603.00729.
  • Qwen Team (2026). Qwen3.5-Omni Technical Report. arXiv:2604.15804.
  • Zuo, Yuxin; et al. (2026). Qwen-AgentWorld: Language World Models for General Agents. arXiv:2606.24597.
  • Su, Jianlin; Lu, Yu; Pan, Shengfeng; et al. (2021). RoFormer: Enhanced Transformer with Rotary Position Embedding. arXiv:2104.09864.
  • Peng, Bowen; et al. (2023). YaRN: Efficient Context Window Extension of Large Language Models. arXiv:2309.00071.

References

  1. ↑ 1.0 1.1 Alibaba Cloud Unveils New AI Model to Support Enterprises’ Intelligence Transformation // Alibaba Cloud. 2023-04-11.
  2. ↑ Alibaba unveils open-sourced AI model similar to Meta's Llama 2 // Reuters. 2023-08-03.
  3. ↑ Alibaba opens AI model Tongyi Qianwen to the public // Reuters. 2023-09-13.
  4. ↑ 4.0 4.1 The Open Source AI Definition — 1.0 // Open Source Initiative.
  5. ↑ 5.0 5.1 5.2 Supported Models and Capabilities Overview // Alibaba Cloud Model Studio. Updated 2026-07-02.
  6. ↑ 6.0 6.1 6.2 6.3 6.4 6.5 Image and Video Understanding // Alibaba Cloud Model Studio. Updated 2026-06-17.
  7. ↑ 7.0 7.1 7.2 7.3 7.4 7.5 7.6 7.7 Qwen3.6 // GitHub.
  8. ↑ 8.0 8.1 8.2 8.3 8.4 8.5 8.6 Qwen: The official repository of Qwen // GitHub.
  9. ↑ 9.0 9.1 Bai, Jinze; et al. (2023). Qwen Technical Report. arXiv:2309.16609.
  10. ↑ 10.0 10.1 10.2 10.3 10.4 Hello Qwen2 // Qwen. 2024-06-07.
  11. ↑ 11.00 11.01 11.02 11.03 11.04 11.05 11.06 11.07 11.08 11.09 11.10 Yang, An; et al. (2024). Qwen2 Technical Report. arXiv:2407.10671.
  12. ↑ 12.0 12.1 12.2 12.3 Qwen2.5-LLM: Extending the Boundary of LLMs // Qwen. 2024-09-19.
  13. ↑ 13.0 13.1 13.2 Qwen Team (2025). Qwen2.5 Technical Report. arXiv:2412.15115.
  14. ↑ 14.0 14.1 QwQ-32B: Embracing the Power of Reinforcement Learning // Qwen. 2025-03-05.
  15. ↑ 15.0 15.1 15.2 15.3 15.4 15.5 Yang, An; et al. (2025). Qwen3 Technical Report. arXiv:2505.09388.
  16. ↑ 16.0 16.1 16.2 16.3 16.4 16.5 16.6 Qwen3: Think Deeper, Act Faster // Qwen. 2025-04-29.
  17. ↑ 17.0 17.1 17.2 17.3 17.4 17.5 Qwen3 // GitHub.
  18. ↑ 18.0 18.1 18.2 18.3 Qwen3-Next-80B-A3B // Qwen. 2025-09-11.
  19. ↑ Qwen3.6 and Qwen3.5 repository — release history // GitHub.
  20. ↑ 20.0 20.1 20.2 20.3 Qwen3-Max: Just Scale // Qwen. 2025-09-23.
  21. ↑ Alibaba's Latest Thinking Model Excels at Adaptive Tool Use // Alibaba Cloud. 2026-01-27.
  22. ↑ 22.00 22.01 22.02 22.03 22.04 22.05 22.06 22.07 22.08 22.09 22.10 Qwen3.5: Towards Native Multimodal Agents // Alibaba Cloud. 2026-02-17.
  23. ↑ 23.0 23.1 23.2 Alibaba Unveils Qwen3.6-Plus // Alibaba Cloud. 2026-04-02.
  24. ↑ Qwen3.6-35B-A3B: Agentic Coding Power, Now Open to All // Qwen. 2026-04-16.
  25. ↑ Qwen3.6-27B: Flagship-Level Coding in a 27B Dense Model // Qwen. 2026-04-22.
  26. ↑ 26.0 26.1 26.2 26.3 Qwen3.7: The Agent Frontier // Alibaba Cloud. 2026-05-21.
  27. ↑ 27.0 27.1 27.2 27.3 27.4 Qwen3.7-Plus: Multimodal Agent Intelligence // Alibaba Cloud. 2026-06-03.
  28. ↑ 28.0 28.1 28.2 Entering the Physical AI Era: Introducing the Qwen-Robot Suite // Alibaba Cloud. 2026-06-17.
  29. ↑ 29.0 29.1 29.2 29.3 29.4 Qwen-AgentWorld // GitHub. 2026-06-24.
  30. ↑ 30.0 30.1 30.2 Zuo, Yuxin; et al. (2026). Qwen-AgentWorld: Language World Models for General Agents. arXiv:2606.24597.
  31. ↑ 31.0 31.1 Bai, Jinze; et al. (2025). Qwen2.5-VL: A Versatile Vision-Language Model for Real-World Agent Tasks. arXiv:2502.13923.
  32. ↑ 32.0 32.1 Wang, Wen; et al. (2025). Qwen2.5-Omni: A Streaming End-to-End Multimodal Model. arXiv:2503.20215.
  33. ↑ 33.0 33.1 Qwen3-Coder // GitHub.
  34. ↑ Qwen3-VL // GitHub.
  35. ↑ 35.0 35.1 Qwen3-Omni // GitHub.
  36. ↑ 36.0 36.1 36.2 36.3 36.4 36.5 36.6 Qwen3.6-27B model card // Hugging Face.
  37. ↑ 37.0 37.1 37.2 37.3 37.4 37.5 Qwen3.6-35B-A3B model card // Hugging Face.
  38. ↑ Peng, Bowen; et al. (2023). YaRN: Efficient Context Window Extension of Large Language Models. arXiv:2309.00071.
  39. ↑ Qwen2.5-1M: Deploy Your Own Qwen with Context Length up to 1M Tokens // Qwen. 2025-01-27.
  40. ↑ Deep Thinking // Alibaba Cloud Model Studio. Updated 2026-06-18.
  41. ↑ Qwen Team (2025). Qwen3-VL Technical Report. arXiv:2511.21631.
  42. ↑ Chu, Yunfei; et al. (2024). Qwen2-Audio Technical Report. arXiv:2407.10759.
  43. ↑ Qwen Team (2026). Qwen3.5-Omni Technical Report. arXiv:2604.15804.
  44. ↑ Qwen-Image // GitHub.
  45. ↑ Qwen3-ASR // GitHub.
  46. ↑ Qwen3-TTS // GitHub.
  47. ↑ Qwen3-Embedding and Qwen3-Reranker // GitHub.
  48. ↑ Qwen-Agent // GitHub.
  49. ↑ Qwen Code // GitHub.
  50. ↑ Qwen Team (2026). Qwen3-Coder-Next Technical Report. arXiv:2603.00729.
Retrieved from "https://systems-analysis.ru/eng/index.php?title=Qwen_(Alibaba)&oldid=3125"