Model Release·Open Weights·MoE·Agentic
Qwen3.8-Max Open Weights: First Max-Class Qwen You Can Download
Aug 12, 2026 open-weights drop of Qwen3.8-2.4T-A95B—the first Max-class downloadable Qwen. Custom license; distinct from the multimodal qwen3.8-max API.
At a Glance
Open weights (Qwen3.8-2.4T-A95B) | Hosted API (qwen3.8-max) | |
|---|---|---|
| Org | Qwen / Alibaba | QwenCloud |
| Params | 2.4T total / ~95B active (512 experts; 10 routed + 1 shared) | Same backbone, managed product |
| Modalities | Text-only | Text + image + video |
| Context | 262,144 native (~1.01M extensible) | 1M default |
| Thinking | Required-on | Optional |
| Tools / vision | BYO | Built-in |
| License | Custom Qwen3.8-Max License | QwenCloud ToS |
| Distribution | Hugging Face / ModelScope | Managed API |
Serving: vLLM, SGLang, TokenSpeed, Transformers.
Architecture: Qwen3.5-family hybrid (Gated DeltaNet + MoE + Gated Attention).
Pricing (QwenCloud API)
| Meter | Per 1M tokens |
|---|---|
| Input | $2.00 |
| Output | $6.00 |
| Implicit cache | $0.25 |
| Explicit cache create | $2.50 |
| Explicit cache read | $0.17 |
Flat across the 1M window. Limits: 2M TPM / 15K RPM. Self-host = your GPUs.
Source: https://www.qwencloud.com/models/qwen3.8-max
License (read before commercial ship)
Qwen3.8-Max License (© 2026 Qwen)—not Apache or MIT:
- Keep the copyright notice and permission notice in all copies or substantial portions.
- If your commercial product or service has >100M MAU or >$20M monthly revenue, prominently display the model name in the UI.
- If you (or affiliates) run a Model as a Service or AI Work Assistant business and aggregate revenue exceeds $50M TTM, obtain a separate license from Qwen before commercial use. Internal-use carve-out applies if you do not expose the model, its outputs, or underlying capabilities to third parties.
- Questions:
[email protected]
Selected vendor scores (API / Max card; self-reported)
Label carefully—these are vendor harness numbers vs Qwen3.7-Max, not LLM Stats verified:
| Benchmark | Qwen3.8-Max | Qwen3.7-Max | Note |
|---|---|---|---|
| Terminal Bench 2.1 | 86.6 | 74.5 | Vendor card |
| SWE-bench Pro | 67.7 | 60.6 | Vendor card |
| PaperBench | 93.0 | 64.8 | Vendor card |
| CoWorkBench | 74.8 | 64.6 | In-house |
| IFBench | 82.8 | 79.1 | Vendor card |
| GPQA Diamond | 92.6 | 92.4 | Vendor card |
| MRCR v2 256K 8-needle | 92.9 | 86.7 | Vendor card |
What's New With This Post
- First Max-class open release — downloadable weights at Max scale.
- Self-host / fine-tune / VPC path under the custom Qwen3.8-Max License.
- Clear split: weights = text + forced thinking; API = multimodal + product features.
Still pending: Qwen3.8-27B open weights are not published yet.
When to Use It
Good fit for open weights
- You need Max-class coding / agentic text quality inside your own cluster or VPC
- Fine-tunes, distillation, or offline eval against the full 2.4T checkpoint
- You can run a 2.4T MoE stack (vLLM / SGLang / TokenSpeed) and accept required-on thinking
Prefer the hosted API (qwen3.8-max)
- You need image / video input, optional thinking, or official built-in tools
- You want 1M context as a managed default without extending the native 262K window yourself
- You would rather pay $2/$6 per 1M than operate the MoE infra
Caveats
- Infra bar is real: serving a 2.4T MoE is not a laptop experiment
- Vendor benches use mixed harnesses (some in-house); do not treat as LLM Stats verified
- Commercial use must clear the custom-license MAU / revenue / MaaS thresholds before ship
Sources
- https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B
- https://www.modelscope.cn/models/Qwen/Qwen3.8-2.4T-A95B
- https://www.alibabacloud.com/blog/qwen3-8-max-a-new-bar-for-coding-and-cowork_603421
- https://qwen.ai/blog?id=qwen3.8
- https://www.qwencloud.com/models/qwen3.8-max
Questions
Frequently Asked Questions
August 12, 2026.
- Hugging Face / ModelScope:
Qwen/Qwen3.8-2.4T-A95B. Catalog ids on LLM Stats:qwen3.8-2.4t-a95b(weights) andqwen3.8-max(hosted API). - A hybrid MoE with about 2.4T total and ~95B active parameters (512 experts; 10 routed + 1 shared). Architecture is Qwen3.5-family hybrid: Gated DeltaNet + MoE + Gated Attention.
- Weights are text-only, with thinking required-on, native context 262,144 (~1.01M extensible), and BYO tools/vision. Hosted
qwen3.8-maxadds text + image + video, optional thinking, 1M context by default, and built-in tools under QwenCloud ToS. - Custom Qwen3.8-Max License (© 2026 Qwen)—not Apache or MIT. Keep copyright + permission notice; above 100M MAU or $20M monthly revenue, prominently display the model name; MaaS / AI Work Assistant businesses above $50M TTM aggregate revenue need a separate Qwen license (internal-use carve-out if you do not expose the model, outputs, or capabilities to third parties). Contact
[email protected]. - Flat across 1M context: $2.00 input / $6.00 output per 1M tokens; implicit cache $0.25; explicit cache create $2.50 / read $0.17. Limits: 2M TPM / 15K RPM. Self-host cost is your GPUs.
vLLM, SGLang, TokenSpeed, and Transformers (per the Hugging Face card).
No. Selected scores in this post are from the vendor Max / API card vs Qwen3.7-Max (mixed harnesses; CoWorkBench is in-house). Independent LLM Stats verification is pending.
- Not yet published as of this post. Only the Max-class
Qwen3.8-2.4T-A95Bdrop is live.
Continue Reading