Alibaba releases Qwen3-Next with lower training costs and faster long document processing

80 billion parameters, with about 3 billion active at a time

Alibaba Cloud has released Qwen3-Next, an open source family of artificial intelligence models designed to deliver faster responses at lower cost. Its lead model, Qwen3-Next-80B-A3B, contains 80 billion parameters but uses only about 3 billion for each token it processes. Parameters are the numerical settings a model learns during training; tokens are the pieces of text it reads and generates.

As of September 12, 2025, Alibaba says the new base model required less than one tenth of the training compute used by Qwen3-32B, its dense predecessor released in April. A dense model uses its full network for each token. In the company's tests, the new base model matches or sometimes exceeds Qwen3-32B, while the wider Qwen3-Next series approaches the quality of the much larger Qwen3-235B on many tasks.

Alibaba also reports more than 10 times higher throughput for prompts longer than 32,000 tokens compared with earlier Qwen releases. Throughput measures how much work a system completes in a given time. The models have a context window of 256,000 tokens, meaning the amount of text they can handle within one request, with an extension to 1 million tokens in some settings.

Developers can download the models or use hosted services through major model platforms and Alibaba's cloud. The company presents the release as a way to bring capable AI to enterprises and developers using cloud infrastructure or consumer grade hardware. Its central promise is practical: spend less computing power on training and answering requests while retaining the capacity to work with lengthy documents and difficult questions.

How the release follows April's Qwen3 models

Qwen3-Next introduces a new architecture, the arrangement of components inside the model. Alibaba combines a sparse Mixture of Experts design, hybrid attention, multi token prediction and changes intended to make training more stable. Together, these choices target both the expense of building a model and the cost of running it afterward.

Training is the stage when a model learns from data. Inference is the stage when a trained model responds to users. A reduction in training compute helps the organization building the model; faster inference can help everyone running it, from a developer testing a prototype to a business serving many customers.

• April 2025: Alibaba released Qwen3-32B, the dense model used as a training efficiency comparison for Qwen3-Next.
• September 12, 2025: Qwen3-Next is available as an open source model family, with instruction and reasoning variants and access through model hubs and hosted services.

Alibaba's training comparison is expressed in GPU hours, a measure that combines the number of graphics processors used with how long they run. The company says Qwen3-Next-80B-A3B-Base needed less than 10 percent of the GPU hours used for Qwen3-32B. The stated comparison concerns computing resources, rather than a disclosed monetary training budget.

Why a large model can do less work per token

Qwen3-Next uses a sparse Mixture of Experts, often shortened to MoE. Instead of sending every token through the entire network, the system chooses a small selection of specialized networks called experts. A lightweight gating module decides which experts to use for each token.

The architecture is described as having 512 experts, with 10 active per token in some layers. Those figures describe the expert selection mechanism; the model's broader specification is 80 billion parameters in total and about 3 billion active at each step. Using the stated rounded figures, the active parameters represent about 3.75 percent of the total.

Alibaba's approach keeps a large pool of learned capacity available while limiting the computation performed for an individual token. The original description compares the compute and memory demands of each forward pass, a single processing step through the network, with those of a much smaller dense model.

MoE also adds complexity. Training must keep the experts balanced so that work is distributed properly. Alibaba says it addressed instability that can arise when sparse MoE is combined with reinforcement learning, a training method that uses feedback to improve model behavior. For developers, the attraction is that processing work depends heavily on the active parameters instead of requiring the entire network to compute every answer.

Making long prompts faster to process

Qwen3-Next combines Gated DeltaNet with classic attention in what Alibaba calls a hybrid attention mechanism. Attention allows a model to examine the text in its input and decide which parts matter to one another. Standard attention becomes expensive as the input grows because its computational cost scales with the square of the sequence length.

The hybrid design mixes standard attention with methods that represent context more economically. Alibaba says this reduces slowdowns as prompts become longer. Its reported gain of more than 10 times higher throughput applies to inputs beyond 32,000 tokens, a workload where processing a large amount of text can otherwise dominate the waiting time.

A context window of 256,000 tokens gives developers room to submit documents of book length, substantial code repositories, contracts or transcripts in a single request. In some settings, the context can extend to 1 million tokens. The release description does not specify the conditions required for that extension.

Alibaba also reports gains approaching linear scaling in some operating ranges, including around seven times faster prefill and around four times faster decode at shorter contexts. Prefill is the stage when a model processes the user's input; decode is the stage when it generates the response. Improving both stages can reduce the time spent waiting for a long document to be read and an answer to arrive.

Predicting several tokens and keeping training stable

Multi token prediction is another part of the architecture. It attempts to predict several upcoming tokens at once, with corrections when those predictions are wrong. Paired with speculative decoding, this can reduce the number of expensive processing passes needed to stream an answer.

Speculative decoding speeds generation by proposing text and checking whether it can be accepted. Qwen3-Next supports this approach and integrates with SGLang and vLLM, popular software frameworks used to run AI models. Those integrations give developers a route to using the architecture within existing inference systems.

Alibaba also identifies Zero Centered RMSNorm and other optimization choices as measures that steady training. RMSNorm is a normalization component that helps control the scale of values inside the network. The company's description places these changes alongside its work on reinforcement learning stability.

The intended result is a combination of faster generation, limited active computation and stable training without a large loss of quality. The actual speed experienced by a user still depends on the deployment settings, including the hardware and software framework.

Two variants for everyday tasks and difficult reasoning

Alibaba has released two models developed through further training of the base network. Qwen3-Next-80B-A3B-Instruct is intended for everyday instruction following, conversation and tool use. Tool use means a model can work with software functions or other connected resources as part of completing a task.

In Alibaba's testing, the Instruct variant performs close to the Qwen3-235B flagship on many general tasks. The company also highlights its performance with very long prompts, reaching the stated context window of 256,000 tokens. That combination is aimed at applications where the model must follow directions while working through a large body of information.

Qwen3-Next-80B-A3B-Thinking is tuned for complex reasoning. Alibaba says it exceeds the performance of its own models in the middle of the Qwen3 range and surpasses the closed Gemini 2.5 Flash Thinking model on several benchmarks selected for evaluation.

Alibaba cites a score of 60.8 percent on SuperGPQA, a difficult test set containing questions at graduate level. The score gives developers one measure of reasoning performance. Teams choosing between the two variants also need results from their own work, including coding, mathematics, retrieval augmented generation and planning that requires several steps.

Retrieval augmented generation combines a language model with a system that finds relevant information from documents or a database. For organizations using that approach, the useful comparison is how well each variant finds, uses and explains the information needed for a real request.

What lower compute demands mean for deployment

Alibaba's claims address two different bills: the compute used to train the base model and the resources used to answer requests after deployment. The training claim is less than 10 percent of Qwen3-32B's compute. The inference claim rests on activating about 3 billion parameters per token, together with the attention and generation changes.

The company associates that limited activity with lower GPU memory pressure and energy demand during inference. Its stated aim is a smaller hardware footprint for a given level of service, or more requests completed on the same setup. For an enterprise, either outcome can make a service less expensive to operate.

Alibaba says the series is tuned to run well on consumer grade hardware, allowing advanced prototypes without large clusters. The release description gives no specific consumer hardware configuration or complete deployment cost, so a developer still needs to match the model and chosen settings to the equipment available.

Long prompts are a particular focus. A team working with contracts, transcripts or large code files can submit substantial material and ask for summaries, comparisons or proposed changes. The reported improvements in prefill and decode address both reading that material and producing a response, rather than generation speed alone.

Where developers can get the models

Qwen3-Next is listed on Hugging Face and ModelScope, where developers can obtain model weights, the learned numerical values needed to run the network. Alibaba Cloud Model Studio provides a managed deployment option for users who prefer hosted infrastructure.

The series is also accessible through the NVIDIA API Catalog, and some variants are mirrored on Kaggle for experimentation. Developers can choose between downloading weights for local inference and calling hosted endpoints through an application programming interface, or API. An API lets software send a request to a service and receive its response.

Alibaba encourages developers to review the license terms for each model, especially for commercial use or services handling large volumes of traffic. Open source access gives teams a route to inspect and build on the models, while the applicable license sets the terms for using a particular release.

The release sits alongside Qwen3-ASR-Flash, a speech transcription component that the team says supports 11 major languages. The article's described workflow combines transcription and translation of audio with language model tasks such as preparing study guides or content outlines. Support for SGLang and vLLM also helps developers move from an initial demonstration toward an operating service.

How Qwen3-Next fits Alibaba's AI competition

Alibaba presents Qwen as a fast growing developer ecosystem and says it has become one of the largest open source communities for developers. Publishing models that others can inspect and adapt is part of a broader push by Chinese AI groups to expand adoption and compete with Western rivals on capability.

Alibaba has also previewed Qwen3-Max, a model with more than 1 trillion parameters that recently ranked sixth on the LMArena chatbot leaderboard hosted by researchers. Qwen3-Next addresses a different part of the same effort: providing substantial capability while reducing the computation required for individual requests.

DeepSeek and Meituan are among the other Chinese companies releasing competitive models and services. The contest concerns access and operating costs as well as performance. Making strong models broadly usable can help smaller organizations build applications without relying entirely on large centralized infrastructure.

Advanced chip export limits add pressure in China, while teams in the United States and Europe also face GPU shortages. Designs that keep most parameters inactive during inference create opportunities for smaller laboratories, startups and universities to run capable models on modest servers. Investors are also watching Alibaba's renewed focus on cloud and AI, alongside an improvement in its stock performance during 2025.

Testing the claims against real workloads

Alibaba's benchmark results and speed claims provide starting points for evaluation. Throughput varies with batch size, meaning how many requests are processed together, as well as hardware, the inference framework and quantization. Quantization uses lower precision representations of model values to reduce resource demands.

Accuracy also changes with the subject, the wording of a prompt and the tools available, including code interpreters and retrieval systems. Vendor evaluations use selected test suites. Independent community testing and customer trials can provide a broader picture of where a model works well and where it needs support.

Legal and financial teams could use the long context window to analyze document sets and prepare tailored summaries that retain citations. Customer support operations could bring knowledge bases and lengthy troubleshooting guides into an assistant. Software teams could submit repositories and logs for code review, refactoring and test generation, while researchers could compare methods and results across long papers and datasets.

Governments and regulated industries could use local deployment with monitoring, audit logging and content filters. Across these applications, teams should measure memory use and throughput at the context lengths they actually need, then test tool use, mathematics, coding reliability and behavior under safety constraints. Applications requiring answers grounded in documents should check the evidence trail. Teams needing consistent reasoning should compare the Thinking variant with their current systems.

The Bottom Line

• Alibaba released the open source Qwen3-Next family, led by an 80 billion parameter model with about 3 billion parameters active per token.
• Alibaba says base model training used less than 10 percent of Qwen3-32B's compute while matching or exceeding its quality in tests.
• The company reports more than 10 times higher throughput for prompts longer than 32,000 tokens.
• The context window reaches 256,000 tokens, with an extension to 1 million tokens in some settings.
• Instruct targets everyday tasks and tool use; Thinking targets complex reasoning and scored 60.8 percent on SuperGPQA in company testing.
• Access includes Hugging Face, ModelScope, Alibaba Cloud Model Studio, the NVIDIA API Catalog and some Kaggle mirrors.
• Alibaba targets cloud and consumer grade hardware, with actual performance depending on deployment settings and workload.