Same weights, different surface
Qwen3.8-2.4T-A95B is the open-weight checkpoint of Alibaba's first Qwen-Max-class release. Qwen3.8-Max is the hosted variant of the same family. Max adds vision input (images and PDFs), non-thinking mode (a response without a leading think block), a 1M default context, and built-in tools such as a code interpreter and web access. Qwen3.8-2.4T-A95B is text-only, requires thinking mode on every call, and defaults to 262K native context (extensible to 1M via YaRN).
Capability and price at a glance
Six published provider rates for the open-weight checkpoint are $2.00 input and $6.00 output per 1M tokens. Max is hosted by Alibaba Cloud and lists at variable rates depending on tier; the cheapest published batch rate for Max is below $2.00 / $6.00. The deciding question is whether the extra surface area on Max (vision, non-thinking, tools, 1M default) is worth the per-token premium.
| Dimension | Qwen3.8-2.4T-A95B | Qwen3.8-Max | Notes |
|---|---|---|---|
| Weights | Open (qwen3.8-max license) | Hosted, closed | Same Qwen-Max-class base |
| Vision input | No (text-only) | Yes | Images and PDFs accepted on Max |
| Non-thinking mode | No | Yes | A95B requires reasoning_effort every call |
| Default context | 262K native, 1M extended | 1M | YaRN on A95B; native on Max |
| Input / Output per 1M | $2.00 / $6.00 | Hosted, varies by tier | Six providers at parity on A95B |
| Tools (code interpreter, web) | Not built in | Built in | OneInfer tool routes either model |
| Fine-tuning | Not on serverless providers | Hosted variant only | Vendor-side |
Benchmark rows to read carefully
Hugging Face attaches the same launch-table scores to both repos. Treat the published numbers as belonging to the Qwen3.8-Max family rather than to either checkpoint in isolation. Terminal Bench 2.1 at 86.6, GPQA Diamond at 92.6, SWE-bench Pro at 67.7, FrontierSWE at 73.5, Toolathlon Verified at 72.5. Independent reproduction: Artificial Analysis GPQA 92.7, HLE 43.0, Intelligence Index 58.1, Coding Index 71.8.
When to pick A95B
- You need open weights for self-hosting, on-prem, or air-gapped deployment.
- You specifically want the lowest published per-token rate for a frontier MoE ($2.00 / $6.00 across six providers).
- You are fine with mandatory thinking mode and a 262K native context.
- You are not using images, PDFs, or built-in tools in this workflow.
Ready to test the workflow?
Create account & add creditsWhen to pick Qwen3.8-Max
- You need vision input (images, scanned PDFs, charts).
- You need non-thinking mode for short, latency-sensitive calls.
- You want a 1M-token default context without an extension step.
- You want built-in tools (code interpreter, web access) without a separate integration.
Migration notes
Moving between A95B and Max is a one-line model-string change on OneInfer. The thinking-mode flag is model-specific: a think block produced on Max cannot be replayed into A95B and vice versa. If you cache response prefixes, clear the cache when switching.
Frequently asked questions
Are Qwen3.8-2.4T-A95B and Qwen3.8-Max the same model?
They share the same Qwen-Max-class weights. Max is the hosted variant and adds vision input, non-thinking mode, a 1M default context, and built-in tools. A95B is the open-weight checkpoint: text-only, thinking-mode required, 262K native context extensible to 1M via YaRN.
Which is cheaper, Qwen3.8-2.4T-A95B or Qwen3.8-Max?
A95B is cheaper on a like-for-like per-token basis: six providers price it at $2.00 input and $6.00 output per 1M tokens. Max is hosted by Alibaba Cloud at variable rates depending on tier; the published batch rate is below the A95B rate, but the standard real-time rate is higher.
Where is the canonical Qwen3.8-2.4T-A95B model page?
The canonical model page with current OneInfer pricing, capabilities, and availability is /models/Qwen/Qwen3.8-2.4T-A95B. This comparison is a focused facet of that entity, not a replacement for it.
How should I treat benchmark or price claims on this page?
Check each claim's provenance label and observed date. Vendor-reported and independently verified numbers are shown as separate evidence classes on this hub.
Is Qwen3.8-2.4T-A95B open source?
Yes — weights are released under the qwen3.8-max license. Confirm the license tag against the official Hugging Face repository before self-hosting at scale.
Put Qwen3.8-2.4T-A95B to work
Fund a controlled evaluation, start with a prepared prompt, and measure quality and cost on your own workload.