Alibaba Cloud: Qwen3.8-2.4T-A95B
qwen3.8-2.4t-a95b- Input / Output
- $2.00 / $6.00
- Context
- 992K
- Max output
- 128K
- Modalities
- Chat
- Capabilities
- Tool usePrompt Cache
About Qwen3.8-2.4T-A95B
Qwen3.8-2.4T-A95B is the open-weight release of the Qwen3.8-Max architecture from Alibaba: a text-only mixture-of-experts model with 2.4 trillion total and 95 billion active parameters. It always reasons before answering and exposes a reasoning-effort setting. Compared with the hosted Qwen3.8-Max, it drops the non-thinking mode, in exchange for downloadable weights.
Where it works well
- Open weights with 512 experts, 11 of them active per token, so the 2.4 trillion parameter total costs only 95 billion per step.
- Adjustable reasoning depth through the reasoning_effort setting, with xhigh, medium, and low levels documented on the model card.
- Alibaba's card reports GPQA Diamond 92.6, SWE-bench Pro 67.7, and Terminal Bench 2.1 86.6.
- Supports tool use, JSON schema output, and prompt caching, so it fits agent and pipeline work.
When to choose another model
- It cannot read images or video; it is a text model.
- Thinking cannot be switched off, so quick conversational turns carry reasoning overhead that Qwen3.8-Flash avoids.
- Self-hosting it takes a very large GPU cluster, which is out of reach for most teams.
Getting started
Create API key
Create a key in Console, then use it with every model on the platform.
Send your first request
Copy the example for your language and run it against the endpoint.
Text to textResponses APIPOST/v1/responsesAPI format:curl https://api.tokenlab.sh/v1/responses \ -H "Content-Type: application/json" \ -H "Authorization: Bearer sk-xxx" \ -d '{ "model": "qwen3.8-2.4t-a95b", "input": "Hello!" }'
Pricing
The Verified price applies to Verified, which costs less on most models. The Official price is the model maker's published price and applies to the more reliable Official route. Auto bills the route that completes the request.
cacheMode · implicit
per 1M tokens- Official price
- Input $2.00 / Output $6.00 / Cache read $0.25 / Text Input $2.00 / Text Input · Cache read $0.25 / Image Input $2.00 / Text Output $6.00
- Verified price
- —
- Discount
- —
cacheMode · explicit
per 1M tokens- Official price
- Input $2.00 / Output $6.00 / Cache read $0.17 / Cache write $2.50 / Text Input $2.00 / Text Input · Cache read $0.17 / Image Input $2.00 / Text Output $6.00
- Verified price
- —
- Discount
- —
| Official priceper 1M tokens | Verified priceper 1M tokens | Discount | |
|---|---|---|---|
| cacheMode · implicit | Input $2.00 / Output $6.00 / Cache read $0.25 / Text Input $2.00 / Text Input · Cache read $0.25 / Image Input $2.00 / Text Output $6.00 | — | — |
| cacheMode · explicit | Input $2.00 / Output $6.00 / Cache read $0.17 / Cache write $2.50 / Text Input $2.00 / Text Input · Cache read $0.17 / Image Input $2.00 / Text Output $6.00 | — | — |
Usage & activity
Success rate is the share of requests that completed. Latency is how long a full response takes; P95 means 95% of requests finished within that time.
Usage & availability
Last 24 hours- Requests
- Success rate
- P95 latency
- Total tokens
Data is based on aggregate user requests, excluding status checks.
Open in Console
Open Qwen3.8-2.4T-A95B in Console with a prompt ready to edit or send.
Help me try qwen3.8-2.4t-a95b with a short message at /v1/responses. Show the reply, latency, and cost.
Use cases
Hard coding problems
Give it a multi-service bug or a large refactor where the extra reasoning pays off, and let it plan before touching files.
Scientific and technical analysis
Work through graduate-level questions, derivations, or paper critiques where careful step-by-step reasoning matters more than latency.
Agent backbones with open weights
Use it where policy requires an open-weight model but the agent still needs frontier-scale reasoning and tool calls.
Structured extraction from long text
Request JSON conforming to a schema from long, messy text sources, with reasoning effort set high for ambiguous inputs.
Prompt examples
Derive the time complexity of this scheduler, then find the case where the amortized bound fails and give a counterexample input.
Extract every obligation, deadline, and responsible party from this text as JSON matching the schema I attached.
Plan a migration of this monolith's billing module into a separate service, with the steps, risks, and order of changes.
This model has conditional pricing. Monthly token totals alone cannot produce a reliable estimate; use the detailed pricing for the request specification, cache, and applicable time window.
FAQ
Is Qwen3.8-2.4T-A95B the same as Qwen3.8-Max?
Not exactly. It is the open-weight twin: same generation and architecture scale, but text-only and always in thinking mode. The hosted Qwen3.8-Max adds a non-thinking mode, built-in tools, and a longer default context.
What do 2.4T and A95B mean?
The model has 2.4 trillion parameters in total, but a gating layer sends each token through only 95 billion of them. That mixture-of-experts design gives large-model knowledge at a much smaller per-token compute cost than a dense model of the same size.
Can I turn thinking off?
No. The model card states that thinking cannot be disabled. You can lower the depth with reasoning_effort set to low, medium, or xhigh. If you need non-thinking replies, use Qwen3.8-Max or a smaller Qwen3.8 model.
Does it accept images?
No. It is a text model and does not take image or video input. Use it for code, documents, and reasoning over text, and pick a different model when screenshots are part of the job.
Which license applies to these weights?
The weights are published by Alibaba's Qwen team on Hugging Face; read the license on the model card before commercial redistribution or fine-tuning. Calling it through an API does not require you to host or license the weights yourself.
How much does Qwen3.8-2.4T-A95B cost?
On TokenLab, Qwen3.8-2.4T-A95B costs Input $2.00 / Output $6.00 per 1M tokens. The pricing table above shows the full breakdown. Rates depend on the billing unit, specification, and usage. Compare matching conditions in the model's detailed pricing; a single rate does not determine the total cost.
What are the context window and output limit of Qwen3.8-2.4T-A95B?
Qwen3.8-2.4T-A95B accepts up to 991,808 tokens of context and returns up to 131,072 tokens in one response.
Which endpoint should Qwen3.8-2.4T-A95B use?
Use https://api.tokenlab.sh/v1/responses for Qwen3.8-2.4T-A95B. The request example below shows the matching code shape.
Which operations does Qwen3.8-2.4T-A95B support?
Qwen3.8-2.4T-A95B supports Text to text. Select an operation above to see its endpoint and request example.
Compare Qwen3.8-2.4T-A95B
Sources
Reviewed Oct 2, 2026
