Qwen3.8-Max

// Overview

Qwen3.8-Max is the August 2026 flagship of Alibaba's Qwen-Max line, generally available from August 3, 2026 through Alibaba Cloud Model Studio (also branded QwenCloud). It succeeds the July 2026 Qwen3.8-Max-Preview shown at the World Artificial Intelligence Conference in Shanghai and sits above the text-only Qwen3.7-Max. Alibaba described it at launch as the most capable model in the Qwen family, and it is the first Max-class Qwen whose weights were also published openly. Alibaba later re-post-trained it for coding and agentic office work and shipped that as the Qwen3.8-Max-0902 snapshot on 2 September 2026.

Technically, Qwen3.8-Max is a sparse Mixture-of-Experts model with 2.4 trillion total parameters and 95 billion active per token, a 1M-token context window (roughly 991K tokens of input and up to 131,072 tokens of output), and a maximum reasoning budget of about 262K tokens. It is natively multimodal — text, image, and video input — and exposes a reasoning-effort control with low, medium, and xhigh settings.

API pricing is $2.00 per million input tokens, $6.00 per million output tokens (including thinking tokens), $0.25 per million tokens for implicit cache reads, and $2.50 per million tokens for explicit cache creation. The weights behind it were subsequently published on Hugging Face as Qwen/Qwen3.8-2.4T-A95B (with an FP8 companion) in August 2026 under a custom licence Qwen names "qwen3.8-max"; that model card describes Qwen3.8-Max as "the official version based on Qwen3.8-2.4T-A95B with more features, such as vision input & non-thinking support, 1M context length by default, official built-in tools".

// Benchmarks

1. Terminal-Bench 2.186.6%
2. PaperBench93%
3. OmniDocBench92.1%
4. DeepSWE 1.156.6%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Pricing

Output pricing includes thinking tokens generated at higher reasoning-effort settings. Explicit cache creation is $2.50 per 1M tokens; explicit cache reads are $0.17 per 1M tokens. The 1M-token context is a single flat tier — the rate does not step up at any threshold.

Pricing source ↗

// Strengths

• Frontier multimodal reasoning: natively accepts text, image, and video input at trillion-parameter scale
• Sparse MoE keeps inference cost tractable: 2.4T total parameters, 95B active per token
• Very long context: 1M-token window flat-priced across the whole range, up to ~131K output tokens per request
• Aggressively cheap for a frontier model: $2 input / $6 output per million tokens, with a 90% discount on cached input
• The first Max-class Qwen to open its weights — published in August 2026 as Qwen3.8-2.4T-A95B on Hugging Face
• Reasoning-effort control (low / medium / xhigh) lets developers trade quality against latency and cost per request

// Best for

• Long-horizon coding agents and multi-step engineering workflows that benefit from the 1M-token context
• Multimodal analysis of screenshots, images, and video alongside text prompts
• Enterprise office and productivity automation that combines documents with charts or diagrams
• Cost-sensitive frontier deployments where the $2 / $6 per-MTok rate and cached-input discount matter
• Applications that intend to self-host: prototype on the hosted API, then swap to the open Qwen3.8-2.4T-A95B checkpoint

// How to access

// Qwen-Max — every version

The full lineage of the Qwen-Max line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.