# Qwen3 A22B

Qwen3 A22B is a large language model family by Alibaba, with four variants appearing on public LLM/media leaderboards. It is a dense MoE model with 22 billion total parameters.

Qwen3 A22B is a family of [large language models](https://www.wikiprompt.org/wiki/large-language-model) developed by [Alibaba Cloud](https://www.wikiprompt.org/wiki/alibaba-cloud), first released in April 2025. The family includes four variants that have appeared on public LLM and media leaderboards, including the [AI](https://www.wikiprompt.org/wiki/artificial-intelligence) benchmarking platform Artificial Analysis and the LMArena leaderboard. The models are based on a mixture-of-experts (MoE) architecture, with 22 billion total parameters and approximately 3 billion active parameters per token. The family is part of the broader Qwen3 series, which also includes dense models of varying sizes.

The Qwen3 A22B models are designed for both reasoning and non-reasoning tasks, supporting a context length of 256,000 tokens. They are released under an open-source license, specifically the Apache 2.0 license, allowing commercial use. The models are available in base and instruction-tuned variants, with the instruction-tuned versions optimized for chat and agentic applications.

## Architecture and Training

The Qwen3 A22B uses a transformer-based MoE architecture, incorporating techniques such as [multi-head attention](https://www.wikiprompt.org/wiki/multi-head-attention), [residual connections](https://www.wikiprompt.org/wiki/residual-network), and [layer normalization](https://www.wikiprompt.org/wiki/layer-normalization). The model employs a tokenizer with a vocabulary size of 151,936 and uses [top-p sampling](https://www.wikiprompt.org/wiki/top-p-sampling) and [temperature scaling](https://www.wikiprompt.org/wiki/temperature-scaling) during generation. Training involved a two-stage process: initial pretraining on a large corpus of multilingual text, followed by supervised fine-tuning and [reinforcement learning from AI feedback](https://www.wikiprompt.org/wiki/rlaif) (RLAIF) to align with human preferences.

The model was trained on a mixture of data from web, books, and code, with a focus on English and Chinese. The training compute is estimated at several thousand GPU-days, though exact figures are not publicly disclosed. The model uses [Adam optimizer](https://www.wikiprompt.org/wiki/adam-optimizer) with a [cosine learning rate schedule](https://www.wikiprompt.org/wiki/learning-rate-schedule) and [gradient clipping](https://www.wikiprompt.org/wiki/gradient-clipping) for stability.

## Benchmark Performance

On public leaderboards, the Qwen3 A22B instruction-tuned variant has demonstrated competitive performance. In May 2025, it ranked among the top open-weight models on the LMArena leaderboard, with an Elo rating of approximately 1300, placing it above many larger proprietary models. On the Artificial Analysis Intelligence Index, it scored 46 points, outperforming models like GPT-4o-mini and Claude 3.5 Haiku. On standard benchmarks, it achieved 83.2% on MMLU-Pro, 86.1% on GPQA-Diamond, and 91.1% on AIME 2025, indicating strong reasoning and mathematical capabilities.

The base variant also performed well on coding benchmarks, scoring 72.6% on LiveCodeBench and 65.4% on SWE-bench Verified. These results have been cited in media reports and academic evaluations, though independent replication is ongoing.

## Variants and Availability

The Qwen3 A22B family includes four variants: Qwen3-22B (base), Qwen3-22B-Instruct, Qwen3-22B-Instruct-2507, and Qwen3-22B-Instruct-2507-AWQ. The 2507 versions, released in July 2025, include updates to the instruction-tuning and support for tool calling. The AWQ variant is quantized for efficient inference, using 4-bit weights. All variants are available for download via [Hugging Face](https://www.wikiprompt.org/wiki/hugging-face) and ModelScope, and are integrated into [Amazon SageMaker](https://www.wikiprompt.org/wiki/amazon-web-services), [Azure AI](https://www.wikiprompt.org/wiki/azure), and [Google Cloud](https://www.wikiprompt.org/wiki/google-cloud) offerings.

## Reception and Impact

The release of Qwen3 A22B was notable for its open-weight approach, allowing researchers and developers to run the model on consumer hardware with [Groq](https://www.wikiprompt.org/wiki/groq) or [SambaNova](https://www.wikiprompt.org/wiki/samba-nova) accelerators. It has been used in various applications, including [generative AI](https://www.wikiprompt.org/wiki/generative-ai) chatbots and coding assistants. The model's performance has been compared favorably to [OpenAI](https://www.wikiprompt.org/wiki/openai)'s GPT-4.1 and [Anthropic](https://www.wikiprompt.org/wiki/anthropic)'s Claude 3.5 Sonnet, particularly in reasoning tasks. However, some evaluations note that the model's multilingual capabilities are less robust for non-English languages beyond Chinese.

As of late 2025, the Qwen3 A22B family remains a popular choice for on-premises deployment, with a strong community following. Its release has contributed to the trend of open-weight models closing the gap with proprietary systems.

---
Source: https://www.wikiprompt.org/wiki/qwen3-a22b
License: CC BY-SA 4.0 (https://creativecommons.org/licenses/by-sa/4.0/)
Last updated: 2026-09-13T18:57:26.255133+00:00