What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Alibaba unveiled Qwen3.5 on February 15, 2026, with an announcement from Alibaba Group dated February 16. The launch centered on Qwen3.5-397B-A17B, an open-weight model designed to handle text, images, video, tool calls, and visual-agent tasks. “Agentic” means it can help plan and carry out steps through connected tools; it does not mean the model can safely run a computer or make consequential decisions on its own. For developers and businesses weighing the cost, the main choice is between using a hosted API and operating a very large model yourself.
What Alibaba launched
The first Qwen3.5 model was Qwen3.5-397B-A17B: 397 billion total parameters, with about 17 billion active in a forward pass. Alibaba described it as a native vision-language model built for reasoning, coding, multimodal understanding, GUI interaction, and agentic tasks. The Qwen announcement is dated February 15, 2026; Alibaba Group’s English announcement is dated February 16. Qwen’s launch announcement and Alibaba Group’s announcement cover the launch.
Model names can be confusing because the original announcement also used the Qwen3.5-Plus name, while Alibaba Cloud’s catalog lists hosted IDs such as qwen3.5-plus separately from qwen3.5-397b-a17b. The current catalog also lists later Qwen3.5-family options. Those listings should not be mistaken for models all released in February.
| Model or label | What it refers to |
|---|---|
| Qwen3.5-397B-A17B | The initial open-weight model announced at launch; 397B total parameters and about 17B active. |
| Qwen3.5-Plus | A launch-era family label also used in connection with the initial release; Alibaba Cloud currently lists a hosted model ID, qwen3.5-plus. |
| Other listed Qwen3.5 options | The cloud catalog includes qwen3.5-122b-a10b, qwen3.5-27b, qwen3.5-35b-a3b, qwen3.5-flash, and qwen3.5-plus. A catalog listing alone does not establish each model’s launch date. |
Alibaba provides open-weight access through Hugging Face, GitHub, and ModelScope, and also offers access through Qwen Chat. Open-weight is the precise term here: availability of weights does not by itself establish that training data, training code, or the full process is open.
#1 Best Overall
What “agentic” means in practice
Qwen3.5 can contribute to an agent workflow by interpreting context, selecting a tool, requesting or taking an action through a connected system, and responding to what happens next. Alibaba describes visual agents that can interact with smartphone and computer interfaces. The model still needs an application or controller to provide tools, credentials, permissions, and a way to observe results. Alibaba Cloud’s model documentation lists function calling and related capabilities.
- Perceive: Read text, images, video, or a screenshot.
- Reason: Break a request into steps and decide what information or action is needed.
- Select a tool: Choose from functions the surrounding application makes available, such as search or an API.
- Act: Return a structured function request or propose an interface action for a GUI controller.
- Observe and iterate: Process the tool result or changed screen, then continue, revise, or stop.
Function calling is a structured request for an application-defined function; it is not the model independently accessing that service. GUI use involves interpreting an interface and working through a controller that can execute actions. An autonomous agent is the broader system that loops through planning and actions toward a goal. A model’s support for tool calls does not guarantee that a long task will finish correctly.
Architecture, context, and what the numbers do—and do not—mean
Qwen3.5-397B-A17B combines sparse mixture-of-experts routing with a linear-attention component using Gated Delta Networks, according to Qwen’s announcement. In a sparse mixture-of-experts model, only some of the model’s parameters are activated for a given token. That helps explain the 17B active-parameter figure, but it does not mean the full 397B checkpoint can run on hardware sized for an ordinary 17B model. Weights, precision, memory overhead, key-value cache, visual tokens, batching, and serving software all affect deployment needs.
Alibaba Cloud lists a 262,144-token context window and a maximum output length of 65,536 tokens for qwen3.5-397b-a17b. Its documentation also lists a maximum input of 258,048 tokens and a maximum chain-of-thought length of 81,920 tokens in thinking mode. These are different limits for different parts of a request, not a guarantee that every deployment exposes identical settings. The model specification is the place to check details for that ID.
Do these 3 things before closing this tab:
1Scan for outdated or missing drivers - takes under a minute2Clear out junk files and repair common Windows errors3Fix the driver behind crashes, sound loss and screen glitchesThe separate hosted model qwen3.5-flash is listed with a one-million-token context window. That figure applies to that variant, not automatically to the 397B model or every Qwen3.5 deployment. Long contexts may add latency and cost, and sheer input length does not ensure that all details receive equal attention. Alibaba’s Flash documentation describes its limits.
Alibaba says language and dialect coverage expanded from 119 to 201. It also presents the architecture as more efficient. Sparse activation and linear attention can reduce some computation, but neither establishes a particular local hardware bill or total cost of ownership. For that, buyers need to account for their workload and serving setup.
What it can be used for
- Visual computer-use workflows: Read a screenshot, locate a field or button, and work through a repetitive interface task when paired with a controller. A visual sketch can also serve as input for generating front-end code.
- Video analysis: Summarize footage, identify events or chapters, or answer questions about what happens over time. Alibaba Cloud’s visual-understanding documentation says its stack supports video input up to roughly two hours; limits can depend on the deployment and input method. See the visual-understanding documentation.
- Coding and software assistance: Generate code from a specification or screenshot, call debugging tools, and organize multi-step development work.
- Enterprise tasks: Combine document or image analysis with search, structured output, customer support, back-office workflows, or field-service assistance.
These are capability categories, not evidence that every task is dependable enough for production without evaluation and safeguards.
How Qwen3.5 differs from Qwen3
Qwen3’s launch emphasized hybrid thinking and non-thinking modes across language models, including dense and mixture-of-experts variants. Qwen3.5’s defining shift was a native multimodal design aimed at agents that can work with images, video, and interfaces as well as language. Alibaba’s Qwen3 announcement is available at the Qwen3 launch page.
| Area | Qwen3 | Qwen3.5 |
|---|---|---|
| Launch emphasis | Hybrid reasoning and open-weight language models | Native multimodal agents |
| Modalities | Primarily language-focused, depending on variant | Text, image, and video input; text output for the documented 397B model |
| Agent focus | Tool use and coding capabilities | Tool use plus visual understanding and GUI interaction |
| Largest launch model | Qwen3-235B-A22B | Qwen3.5-397B-A17B |
| Active parameters | About 22B for Qwen3-235B-A22B | About 17B for Qwen3.5-397B-A17B |
| Context | Varies by model | 262,144 tokens for the 397B model; one million for hosted Flash |
| Architecture | Dense and MoE variants | Hybrid linear attention and sparse MoE design |
Fewer active parameters do not automatically make Qwen3.5 faster or cheaper in a real deployment. Hardware, quantization, request length, visual input, batching, and serving infrastructure all matter.
Hosted API or open-weight deployment?
Use Alibaba Cloud Model Studio for a hosted API
Model Studio is the lower-friction route if you want to call a hosted model rather than manage a large checkpoint. Alibaba offers official Qwen APIs and OpenAI-compatible APIs; endpoint, model, and feature availability differ by region. Start at the Model Studio overview and confirm the model ID and regional endpoint in your account.
- Create or access an Alibaba Cloud account and open Model Studio.
- Create a workspace and API key.
- Select the deployment region and copy that region’s endpoint.
- Choose an available model ID, such as
qwen3.5-397b-a17b, and configure tools or structured outputs in your application if needed.
Alibaba’s OpenAI-compatible pattern looks like this; replace the base URL with the endpoint for your workspace and region:
from openai import OpenAI
client = OpenAI(
api_key="YOUR_ALIBABA_CLOUD_API_KEY",
base_url="https://YOUR_WORKSPACE_ID.maas.aliyuncs.com/compatible-mode/v1"
)
response = client.chat.completions.create(
model="qwen3.5-397b-a17b",
messages=[
{"role": "user", "content": "Analyze this document and return structured findings."}
]
)
print(response.choices[0].message.content)The sample illustrates the API pattern, not a universal endpoint: base URLs, supported models, keys, and features vary across regions. A code snippet configured for one region may not work unchanged in another.
Understand the API pricing signal
Alibaba Cloud’s documentation lists these standard rates for qwen3.5-397b-a17b in the United States/Virginia and Germany/Frankfurt global scopes. The documentation notes that standard prices may exclude promotions. This is token-based hosted API pricing, not the cost of running the open weights.
| Input length tier | Input rate | Output rate |
|---|---|---|
| Up to 128k input tokens | $0.172 per 1 million input tokens | $1.032 per 1 million output tokens |
| Above 128k and up to 256k input tokens | $0.43 per 1 million input tokens | $2.58 per 1 million output tokens |
For qwen3.5-flash, Alibaba lists a one-million-token context window and standard rates beginning at $0.029 per 1 million input tokens and $0.287 per 1 million output tokens in its China/Beijing documentation table. Those China figures should not be applied to a U.S. account; check the model’s region-specific pricing. Sources: 397B pricing and specification and Flash pricing and specification.
Download the weights only with a deployment plan
Open-weight deployment offers more control over hardware, data location, inference stack, adaptation, and private networking. The trade-off is operational: the 397B checkpoint is not a lightweight download for a typical workstation. Multi-GPU infrastructure or substantial GPU memory may be required; quantization can lower memory needs but may affect quality or compatibility. Compute, storage, engineering, monitoring, and support can all carry costs even when weights are downloadable.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.How strong is it—and what should buyers test?
Alibaba reports strong results across language, coding, reasoning, agent, and multimodal benchmarks. Treat those as vendor-reported results rather than a universal ranking: performance depends on the exact model version, benchmark version, prompts, tools, and competitor versions used. A result on selected tests does not prove that Qwen3.5 is better overall than GPT, Claude, Gemini, DeepSeek, or another model.
For a purchase or deployment decision, test representative tasks from your own workload: tool selection, visual accuracy, recovery from errors, output format compliance, latency, and cost. Compare models under the same instructions and tool access, and check the failure rate on actions that matter rather than judging only polished demonstrations.
Risks to consider before giving an agent access
A model connected to tools can choose the wrong function, misread a screen, hallucinate that an action succeeded, repeat an action, or expose sensitive data. It may also follow malicious instructions embedded in a webpage or document. In a computer-use setting, mistakes could affect email, customer records, payment systems, administrative consoles, production infrastructure, or irreversible changes.
- Limit tools with allowlists and least-privilege credentials.
- Use sandboxing, rate limits, retry limits, audit logs, and clear stop conditions.
- Require confirmation or human approval for payments, deletion, deployment, or other consequential actions.
- Keep authentication boundaries separate from model prompts, and treat web pages and documents as untrusted input.
- Evaluate the full application loop, not just the model’s text response.
Is Qwen3.5 still a sensible choice?
Qwen3.5 remains relevant if a specific workflow benefits from its multimodal or visual-agent capabilities, you need its open weights, or you are maintaining a compatible deployment. But it should not be described as Alibaba’s newest Qwen generation: Alibaba Cloud’s model catalog, as documented on August 18, 2026, lists newer Qwen3.6 and Qwen3.7 models among the options for new projects. Check the current catalog before committing to a model. Alibaba Cloud’s model list identifies the available generations.
Quick Recap
- Choose hosted Model Studio when you want to prototype without managing inference infrastructure, after confirming regional availability and pricing.
- Consider a smaller Qwen3.5 variant if the largest checkpoint is excessive for the task, latency target, or budget.
- Evaluate newer Qwen models for a new Alibaba-based production project where current recommendations and support matter.
- Compare other providers if enterprise controls, support, compliance requirements, or your own task evaluations favor another service. Model Studio’s catalog includes alternatives such as DeepSeek, Kimi, GLM, and MiniMax; availability depends on region. See the Model Studio overview.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
