Understand Qwen3.8 Max before you size the hardware.
Qwen3.8 Max is Alibaba’s flagship model, and Qwen3.8 is the first generation in which the Max class was released as open weights at all: the hosted Max builds on the published 2.4-trillion-parameter checkpoint. Its architecture is a 3-to-1 hybrid — three linear-attention layers for every sparse-attention layer — which is what makes a 262,144-token context affordable to serve.
Architecture, benchmarks and licence terms on this page come from Qwen’s own model cards and the Qwen3.8 release blog. Qwen publishes no per-token price on the model card; hosted rates live on Qwen Cloud.
The architecture Qwen draws for the 3.8 generation
Qwen’s cards carry one static figure: the block diagram of the hybrid attention stack. The 2.4T card describes that layout in text, and the diagram comes from the sibling Qwen3.8-Flash-Next card, which draws it out.
The diagram belongs to Qwen and is reproduced with credit: Qwen3.8 Flash Next model card
How Qwen describes Qwen3.8 Max
Qwen calls Qwen3.8 the most capable generation in its open-model family, and the release blog is titled “A New Bar for Coding and Cowork”. The architecture underneath it is the more interesting part: a deliberate mix of linear and sparse attention.
The Max class went open
Qwen says this is the first time a Qwen-Max-class model has been released openly, and publishes it as Qwen3.8-2.4T-A95B. The hosted Qwen3.8 Max is that checkpoint plus serving features the weights do not carry.
Three linear layers to one sparse layer
The hidden layout is 23 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE)) across 92 layers. Gated DeltaNet carries 128 linear-attention heads for values against 16 for queries and keys; the sparse path uses 64 query heads against 4 key-value heads with a 256-wide head dimension. The linear path is what makes the long context affordable.
A very large vocabulary and an N-gram embedding layer
The token embedding is 248,320 entries, padded. The 2.4T card also carries multi-token prediction trained over multiple steps, and the generation’s block diagram shows an N-gram embedding layer feeding layer two.
What Max adds over the raw weights
Qwen’s card is explicit: the hosted Max adds vision input, non-thinking mode, a 1M context window by default and official built-in tools. The published weights are text-only and require thinking mode — the card states that thinking cannot be turned off there.
Benchmarked on an unusually wide surface
The comparison table covers coding agents, general agent tasks, professional work in law, finance and health, and long-context retrieval. That breadth is the argument: Qwen is selling a general worker rather than a coding model, and the table shows where it still loses.
What the model card reports
Qwen’s own comparison table, with the best rival column added so the rows where Qwen3.8 Max loses stay visible. It is a vendor-run table, and the competitor figures are those vendors’ published scores.
| Benchmark | Qwen3.8 Max | Best rival in the same table |
|---|---|---|
| Terminal Bench 2.1 | 86.6 | GPT 5.6 Sol — 88.8 |
| SWE-bench Pro | 67.7 | Fable 5 — 80.0 |
| PaperBench | 93.0 | GPT 5.6 Sol — 90.5 |
| CoWorkBench | 74.8 | Fable 5 — 75.9 |
| WideSearch | 81.9 | Fable 5 — 81.2 |
| IFBench | 82.8 | GPT 5.6 Sol — 72.7 |
| GPQA Diamond | 92.6 | GPT 5.6 Sol — 94.1 |
| HLE | 43.6 | Fable 5 — 53.3 |
| HLE with tools | 56.2 | Fable 5 — 64.5 |
| MRCR v2 256K | 92.9 | GPT 5.6 Sol — 93.8 |
| HealthBench | 60.2 | GPT 5.6 Sol — 55.3 |
Qwen ran its own Terminal Bench 2.1 evaluation with the Claude Code harness at average@10, while taking other models’ best published scores from different harnesses — Artificial Analysis for the Claude models, Codex for GPT-5.6 Sol. The card flags that Fable 5 results may involve fallbacks, and that several of its rows are in-house benchmarks a reader cannot reproduce.
Documented specifications
Every row below is stated on Qwen’s model card for the 2.4T checkpoint.
- Developer
- Alibaba Qwen Team
- Model
- Qwen3.8 Max
- Open weights
- Qwen3.8-2.4T-A95B
- Parameters
- 2.4T total, 95B activated
- Layers
- 92
- Hidden layout
- 23 × (3 × (Gated DeltaNet → MoE) → 1 × (Gated Attention → MoE))
- Experts
- 512 total, 10 routed + 1 shared per token
- Vocabulary
- 248,320, padded
- Context
- 262,144 native, extensible to 1,010,000
- Max adds
- Vision input, non-thinking mode, 1M context, built-in tools
- Reasoning effort
- xhigh (default), medium, low
- Licence
- Qwen3.8-Max licence
Documented access channels
There are two ways in, and they are not the same model: the downloadable 2.4T checkpoint, or the hosted Max that adds vision, a non-thinking mode and built-in tools.
Official model card
Qwen’s own card for the 2.4T checkpoint: the architecture, the context length, the sampling parameters and the full comparison table.
OpenQwen3.8 Max on Qwen Cloud
The hosted model page, which lists the features the checkpoint does not carry: vision input, non-thinking mode, 1M context by default and the official built-in tools.
OpenQwen Studio
The consumer entry point, which can be pointed straight at Qwen3.8 Max for a hands-on check before you commit to an integration.
OpenServing frameworks
Qwen links official recipes for SGLang, vLLM and TokenSpeed and recommends them for production or high-throughput work, warning that throughput varies significantly between frameworks and versions.
Qwen3.8 Max questions
Is Qwen3.8 Max open weights?
Partly, and the split matters. Qwen released the underlying 2.4T checkpoint openly — the first Max-class model to get that treatment — but the hosted Qwen3.8 Max adds vision input, non-thinking mode, a 1M context window by default and official built-in tools that the checkpoint does not have. The licence is the custom qwen3.8-max licence, not an OSI licence.
Can I run it on my own hardware?
This is a 2.4T-parameter model with 95B activated per token, so it is data-centre territory rather than workstation territory. Qwen documents serving through SGLang, vLLM and TokenSpeed and recommends the latest framework versions; it publishes no minimum hardware figure.
How large is the context window?
262,144 tokens natively, extensible to 1,010,000. The hosted Max runs with a 1M context by default. On Qwen’s own MRCR v2 256K retrieval test the model scores 92.9, against 93.8 for GPT 5.6 Sol and 86.7 for its own predecessor, Qwen3.7-Max.
Why does the model card say thinking cannot be disabled?
Because the published checkpoint is text-only and requires thinking mode for every interaction: multimodal inputs are unsupported there, and every response begins with reasoning before the final answer. Non-thinking support is one of the features Qwen reserves for the hosted Max.
How do I control how much it thinks?
Through reasoning_effort, which takes xhigh (the default), medium or low. A second switch, preserve_thinking, is on by default and retains reasoning context from earlier messages. Qwen recommends temperature 1.0, top_p 0.95, top_k 20, min_p 0.0, presence_penalty 0.0 and repetition_penalty 1.0.
How much does Qwen3.8 Max cost?
Qwen’s model card publishes no per-token rate, so this page states none. The card points to Qwen Cloud for managed inference, and its model page is the place to check current pricing.
Is it actually better than the closed frontier models?
It depends on the row. On Qwen’s own table Qwen3.8 Max leads PaperBench at 93.0 against 90.5, WideSearch and IFBench, and trails on SWE-bench Pro at 67.7 against 80.0, HLE at 43.6 against 53.3 and Terminal Bench 2.1 at 86.6 against 88.8. Several rows are Qwen’s in-house benchmarks, and the card notes that different models were evaluated with different harnesses.