Alibaba's Qwen3.5 Offers Local AI and Low API Prices, With Mixed Results on Complex Coding

Three Open Models Put Local AI in Focus

Alibaba's Qwen team has released three downloadable Qwen3.5 Medium models under the Apache 2.0 license, alongside a proprietary cloud version whose advertised prices start at $0.10 per million input tokens and $0.40 per million output tokens. As of February 27, 2026, the release sharpens competition over how much capable artificial intelligence should cost and where it should run.

The downloadable models are Qwen3.5-35B-A3B, Qwen3.5-122B-A10B and Qwen3.5-27B. All are available on Hugging Face and ModelScope. Their permissive license allows businesses and independent developers to download, modify and deploy them commercially without licensing fees. Qwen3.5-Flash, the fourth variant, is available exclusively through the Alibaba Cloud Model Studio API.

Alibaba says Qwen3.5-35B-A3B surpasses OpenAI's GPT-5-mini and Anthropic's Claude Sonnet 4.5 on evaluations covering knowledge retrieval and visual reasoning. Sonnet 4.5 launched about five months earlier, making the claimed gap closing rapid in an industry where development cycles typically span quarters or years.

The attraction extends beyond benchmark scores. Alibaba presents the models as a way to bring advanced AI onto locally controlled hardware, using selective computation and compression to reduce the resources needed. Independent coding tests, however, show a less consistent picture, particularly when a model must coordinate changes across many files.

For businesses choosing between a hosted service and their own equipment, Qwen3.5 brings three decisions together: the quality required for a task, the cost of running it and whether sensitive information can leave the organisation's infrastructure.

How the Architecture Reduces Computation

Qwen3.5 combines Gated Delta Networks with a sparse Mixture-of-Experts architecture, usually shortened to MoE. The practical purpose of this hybrid design is to process information efficiently while drawing on a much larger collection of model parameters.

Parameters are the numerical values a model learns during training. A traditional dense model activates all its parameters for each query. An MoE model instead routes work to selected groups of parameters, much as a team might send a question to the specialists best placed to answer it.

Qwen3.5-35B-A3B contains 35 billion parameters but activates about 3 billion for each token it processes. Tokens are the pieces of text a model reads or generates. The model uses 256 expert modules, with eight routed experts and one shared expert handling particular aspects of the work.

That selective activation reduces the computation required during inference, the stage when a trained model produces answers. Alibaba says the architecture cuts response delays while maintaining output quality. The difference between total and active parameters is central to the model's appeal: it holds a large collection of learned information while using only part of it at each processing step.

Alibaba also describes the model as retaining nearly all its accuracy when its weights are compressed to four bits through quantization. Quantization stores the numerical values that make up a model with fewer bits, reducing memory requirements. Models can lose quality during this process, so resilience under compression matters for deployment on smaller machines.

The company's claim is especially relevant to developers seeking local AI. A model that requires less computation and tolerates substantial compression offers more flexibility over the equipment on which it can operate.

What Local Hardware Can Support

Alibaba presents Qwen3.5-35B-A3B as capable of handling a context window exceeding 1 million tokens on consumer GPUs with 32GB of video memory through four bit quantization. A context window is the amount of information a model can work with during an interaction, including material supplied to it and the response it generates.

A GPU with 32GB of video memory is still expensive consumer equipment. Even so, the claimed threshold puts this deployment option within reach of more developers than the server infrastructure generally associated with comparable AI capabilities. Video memory, often called VRAM, holds the information a GPU needs while performing its work.

The other models target different equipment and workloads. Qwen3.5-27B is positioned for efficient deployment and supports context exceeding 800,000 tokens. Qwen3.5-122B-A10B targets more demanding use and requires server GPUs with 80GB of video memory for maximum performance.

Alibaba's tiered approach allows organisations to choose a model around their existing infrastructure. A developer working with a single GPU, a business operating powerful servers and a customer using a hosted API have different options within the same family.

Long context windows also broaden the material a model can process. The release describes uses involving large document collections, video lasting an hour and extensive institutional datasets. For a business, the appeal is the ability to work with substantial internal material on equipment it controls.

Early feedback on Hugging Face includes enthusiasm for the models' ability to use tools and remain stable after quantization. Developers have reported successful deployments on consumer hardware previously considered insufficient for AI of this capability. Those experiences support interest in the efficiency claims, while the independent coding results show why model selection still depends on the work involved.

Knowledge Tests and Coding Tests Tell Different Stories

Alibaba's published comparisons place Qwen3.5-35B-A3B ahead of GPT-5-mini and Claude Sonnet 4.5 on measures including MMMLU, which tests multilingual knowledge, and MMMU-Pro, which tests visual reasoning. These company comparisons using third party evaluations underpin the claim that a smaller locally deployable model can compete with leading proprietary alternatives.

Independent SWE-rebench testing points to a different competitive picture in software development. That evaluation uses bug fixes involving real GitHub pull requests rather than synthetic benchmark problems. Claude Sonnet 4.5 achieved the highest pass rate, at 55.1%.

Anthropic's model also solved several complex instances that none of the other tested models resolved. Those included problems in the python-trio, cubed-dev and canopen repositories. Its lead in that setting matters for users who need AI to repair working software, rather than answer a contained knowledge question.

The chronology helps explain the attention surrounding Alibaba's release:

• About five months before the February 27 article, Anthropic launched Claude Sonnet 4.5.
• In the days before February 27, Alibaba unveiled the Qwen3.5 Medium series.
• By February 27, the three Apache 2.0 models were available on Hugging Face and ModelScope, while Flash was offered through Alibaba Cloud Model Studio.

The central choice for users is therefore more specific than which model leads a general ranking. Knowledge retrieval, visual reasoning and software repair ask a model to perform different kinds of work, and Qwen3.5's results vary across them.

Why the Smaller Dense Model Did Better on Some Tasks

Independent APEX testing examined 70 tasks involving actual codebases and real problems. Its results exposed differences within the Qwen3.5 family, including cases where the smaller dense model outperformed a model using the more computationally selective MoE design.

Qwen3.5-27B scored 1384 Elo and outperformed DeepSeek V3.2. Elo is a comparative rating used to rank performance within an evaluation. The result made the 27B model a credible option for deployment on a single GPU.

Qwen3.5-35B-A3B scored 1256 Elo on agentic workloads, below the 27B model. Agentic work involves taking a sequence of actions toward a goal, rather than producing a single answer. In software development, that can mean inspecting files, making changes and following through on the consequences across a project.

The much larger Qwen3.5-397B configuration also struggled as task complexity rose. Its rating fell from approximately 1550 Elo on hard tasks to 1194 on tasks at the master difficulty level, where success required coordinating many files over extended reasoning chains.

The researcher behind the independent analysis, who was not named and had spent more than $3,000 on compute for testing, described the difficulty this way:

When it needs to coordinate across many files over many steps, it just loses track of what it's doing.

The APEX results suggest that Qwen3.5 can perform well on contained problems while encountering difficulties on extended coding assignments. Claude Sonnet 4.5 showed particular strength in complex agentic work. The dense 27B model's stronger showing also gives developers a reason to test several variants against their own tasks rather than choose by parameter count alone.

Thinking Mode and Tools Extend the Models' Role

All Qwen3.5 models ship with Thinking Mode enabled by default. Before producing their final output, they generate reasoning chains separated by specific tags, working through a problem before presenting an answer.

The approach follows chain-of-thought methods that have improved accuracy in other leading models. For a reader unfamiliar with the term, the basic idea is to give the model room to work through intermediate steps instead of immediately producing a conclusion.

The series also supports tool calling. This allows a model to interact with external systems, search the web and execute code. Those capabilities expand its role from generating text to carrying out steps within a broader workflow.

Qwen3.5-Flash attaches separate charges to some tools. Web search costs $10 per 1,000 calls, while code interpreter access is currently offered without charge for a limited time. Alibaba has not given an end date for that temporary free access.

Tool use makes the coding evaluations especially relevant. A model may have access to the tools needed for a task while still finding it difficult to keep track of a long sequence of decisions. Businesses considering automated workflows therefore have both the tool costs and the model's ability to complete the work to assess.

How Flash Pricing Compares

Qwen3.5-Flash is priced at $0.10 per million input tokens and $0.40 per million output tokens. Input tokens are the material sent to a model; output tokens are the material it generates. The separate prices matter because a service processing long documents can have a different spending pattern from one producing lengthy answers.

Claude Sonnet 4.5 costs $3.00 per million input tokens and $15.00 per million output tokens. GPT-5.2 costs $1.75 and $14.00 respectively. GLM-5 from Z.ai, another alternative for users interested in locally deployable model families, costs $1.00 and $3.20 through its hosted service.

Using those figures, Sonnet 4.5's input rate is 30 times Flash's rate, while its output rate is 37.5 times as high. GPT-5.2's rates are 17.5 times as high for input and 35 times as high for output. GLM-5's corresponding rates are 10 times and eight times as high.

The release's broader pricing pitch describes a difference of 30 to 100 times against Western competitors. The listed prices give customers a more specific comparison for the named services and show why the size of any saving depends on the balance between incoming and generated text.

For businesses handling large volumes, those differences can create a strong incentive to consider Flash. A lower token price has to be weighed alongside the model's performance on the intended task and charges for tools such as web search.

Local Deployment Changes Control Over Data

Finance, healthcare and government contracting are among the sectors where organisations face requirements restricting the transfer of sensitive information to external cloud AI services. Running a model within private infrastructure gives those organisations an option beyond sending each request to a proprietary API.

Qwen3.5's downloadable weights allow businesses to process internal documents and institutional datasets on systems they control. The large context windows are part of that appeal, making it possible to work with substantial collections of information without sending the material to an external AI endpoint.

The Apache 2.0 license also permits modification and commercial deployment. Organisations can fine tune the models for specialist vocabulary, internal knowledge and particular workflows. That flexibility is useful where technical terminology and regulatory documentation shape everyday work.

Alibaba has also released the base model alongside instruction tuned variants. A base model gives researchers and developers a starting point for further training and architectural study, while instruction tuned models are adapted to respond to requests. This access contrasts with proprietary services that expose a model through an API without providing its weights.

Local deployment and Flash nevertheless represent different data arrangements. Flash operates through Alibaba Cloud Model Studio. The downloadable models can be deployed in an air gapped environment, meaning infrastructure isolated from external networks, which reduces many concerns associated with sending information to a hosted service.

A Wider Contest Between Open and Closed AI

Qwen3.5 joins a broader group of Chinese AI releases competing with proprietary American systems. Z.ai's GLM-4.5 family is also available under Apache 2.0 and includes features such as automated PowerPoint generation. Moonshot's Kimi K2 and DeepSeek's open models add further choices for developers.

That competition reaches beyond model scores. Downloadable weights, commercial licensing terms, hardware requirements and API prices all affect whether an organisation can afford to adopt a model and adapt it to its work.

Western developers face their own pressures. Meta's Llama 4 family drew criticism from some AI researchers over concerns about benchmark gaming and inconsistent performance in practical use. OpenAI delayed its planned open source frontier model release from July to an unspecified later date, giving competing open models more time to gain developer attention.

Businesses considering Chinese developed models also face questions about potential content restrictions and data governance, particularly when they use hosted APIs. The ability to operate Qwen3.5 locally, including on isolated infrastructure, is therefore a central part of its competitive position.

Alibaba's release combines accessible weights, low hosted prices and several hardware tiers. Its efficiency claims and early developer feedback make it a serious option for local deployment. The independent tests add an equally practical consideration: complex work involving many files and many steps remains an area where performance varies sharply, both between Qwen variants and against Claude Sonnet 4.5.

Key Points

• As of February 27, 2026, Qwen3.5-35B-A3B, Qwen3.5-122B-A10B and Qwen3.5-27B are available under Apache 2.0. Flash remains proprietary.
• Qwen3.5-35B-A3B contains 35 billion parameters and activates about 3 billion per token.
• Alibaba claims context exceeding 1 million tokens on a consumer GPU with 32GB of video memory through four bit quantization.
• Company comparisons show strong knowledge and visual reasoning results, while independent tests reveal weaknesses on complex coding assignments.
• APEX rated the dense 27B model at 1384 Elo, above the 35B MoE model's 1256 on agentic workloads.
• Flash costs $0.10 per million input tokens and $0.40 per million output tokens, with separate charges for web search.
• Thinking Mode is enabled by default, and the models support tools for web search, code execution and interaction with external systems.
• Local deployment and modification give organisations more control over sensitive data and specialised workflows.