Overview
Qwen3.8-2.4T-A95B is the downloadable version of Qwen 3.8 Max, the flagship of Alibaba's Qwen 3.8 release. Alibaba announced the Max on August 3, 2026 and published these weights on August 12. It is a sparse mixture of experts: 2.4 trillion parameters in total, of which 95 billion are active for any given token.
| Specification | Qwen3.8-2.4T-A95B |
|---|---|
| Total parameters | 2.4T |
| Active parameters | 95B per token, 10 of 512 experts |
| Architecture | Sparse MoE, hybrid attention (Gated DeltaNet + Gated Attention) |
| Layers | 92 |
| Context window | 262,144 tokens |
| Modalities | Text input, text output |
| Reasoning | Reasons on every request |
| Release date | August 12, 2026 |
| License | Qwen3.8-Max license |
Sparse routing is what makes a model this size servable at all: only 10 of the 512 experts run per token, so the compute per token stays near a 95B model while the weights still have to sit in memory. That memory requirement is the reason this one belongs on a multi-GPU node rather than a workstation. The dense Qwen3.8-27B from the same release is the model to run at home.
Qwen3.8-2.4T-A95B benchmarks
The numbers Alibaba publishes on this repository are scored on Qwen 3.8 Max, the served version of these weights, against its predecessor and three frontier API models:
| Qwen3.8-Max | Qwen3.7-Max | Claude Opus 4.8 | Fable 5 | GPT 5.6 Sol (max) | |
|---|---|---|---|---|---|
Terminal Bench 2.1 Agentic terminal-use tasks | 86.6 | 74.5 | 84.6 | 84.6 | 88.8 |
SWE-bench Pro Real-world software engineering, harder split | 67.7 | 60.6 | 69.2 | 80.0 | 64.6 |
DeepSWE 1.1 Agentic coding | 56.6 | 21.6 | 59.0 | 70.0 | 73.0 |
GPQA Diamond Expert-level science questions | 92.6 | 92.4 | 92.0 | 92.6 | 94.1 |
Humanity's Last Exam Expert-level questions across every domain | 43.6 | 41.4 | 45.7 | 53.3 | 47.2 |
Toolathlon Verified Long-horizon tool use | 72.5 | 49.7 | 76.2 | 77.9 | 74.9 |
WideSearch Broad web research | 81.9 | 75.2 | 72.9 | 81.2 | - |
IFBench Instruction following | 82.8 | 79.1 | 62.2 | 63.5 | 72.7 |
The Max leads its predecessor on every row here, and it takes instruction following and broad web research outright. On agentic coding it sits behind Fable 5 and GPT 5.6 Sol, which is the gap Alibaba is closing with each release rather than one it claims to have closed.
Qwen3.8-2.4T-A95B hardware requirements
The system requirement to check is memory, and at this size it is measured in hundreds of gigabytes. There is no consumer configuration that runs these weights. The smallest GGUF build published for the model is still 397 GB:
| Memory | Build to pick | File size |
|---|---|---|
| 400 GB | UD-Q1_0 | 397 GB |
| 512 GB | UD-IQ1_S | 508 GB |
| 768 GB | UD-IQ2_XS | 731 GB |
| 1 TB and up | UD-IQ3_XXS | 956 GB |
This is the same class of hardware as Kimi K3: a rented multi-GPU node or a very large unified-memory machine, served through vLLM or SGLang. If you want Qwen 3.8 on your own machine, run the 27B instead.
How to run Qwen 3.8 in Atomic Chat
Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.
- Download Atomic Chat for your platform and open it.
- Search for Qwen3.8 in the model browser and open Download Options.
- Pick the build that fits the memory you have, then start a chat.
The full walkthrough, with the quant table and the llama.cpp commands, is in our guide to running Qwen 3.8 locally.
Atomic Chat downloads and runs whatever fits your machine, which for most people means the 27B. See the full lineup of Qwen models to run locally.
License
The weights ship under Alibaba's own Qwen3.8-Max license rather than Apache 2.0, which is what the rest of the Qwen line uses. Read the license file on the model repository before any commercial deployment, because the terms are set by Alibaba and not by a standard open-source licence.