Qwen3.8 Max

Alibaba officially announced Qwen3.8-Max on August 3, 2026 as the most powerful model in its Qwen series to date. The model features 2.4 trillion total parameters with 95 billion active parameters per query, a context window of up to 1 million tokens, and native multimodal support for visual intelligence. It ranks fift Qwen3.8-Max is accessible via APIs on Alibaba Cloud Model Studio for global developers, with model weights scheduled for release the following week. It can also be experienced on QwenWork, Alibaba's all-in-one workplace AI agent platform. The release marks Alibaba's return to open-sourcing its top-tier models after kee

Qwen3.8-Max went live through Alibaba's QwenCloud API around August 2-3, 2026 and was positioned under the headline "A New Bar for Coding and Cowork" on Alibaba's Qwen.ai blog, targeting software engineering and multi-step agent tasks. The open-weight variant, packaged as Qwen3.8-2.4T-A95B, landed on Hugging Face and M By early September 2026, Qwen3.8-Max had become one of the more closely watched releases for engineers building coding agents, reporting an 86.6 score that approaches Opus 5 territory. The release lands amid a crowded September for frontier AI, with OpenAI, Anthropic, and Google shipping new models within days of each

Qwen3.8-Max is a 2.4 trillion total parameter / 95 billion active parameter Mixture-of-Experts model released by Alibaba on August 3, 2026, after a preview teased on July 19, 2026. It uses 512 experts (10 routed plus 1 shared) across 92 layers, adopts a hybrid Gated-DeltaNet and full-attention architecture built on the The licence for Qwen3.8-Max is a custom "qwen3.8-max" agreement close to MIT but with two thresholds: attribution is required above 100 million monthly active users or $20 million monthly revenue, and a separate licence is needed only for Model-as-a-Service or "AI Work Assistant" businesses exceeding $50 million over t

Qwen3.8-Max reached general availability on August 3, 2026 through QwenCloud and Alibaba Cloud Model Studio, following a mid-July preview. The model is a 2.4 trillion parameter Mixture-of-Experts with 95 billion active parameters per query, built on the Qwen3.5 architecture, accepting text, images, and video as input w Alongside the flagship, Alibaba is open-weighting a second checkpoint called Qwen3.8-27B aimed at ordinary on-premise GPU hardware, with weights for both checkpoints scheduled to appear on Hugging Face and ModelScope the following week. This makes Qwen3.8-Max the first model in the Qwen-Max class to have its weights pu

Qwen3.8-Max was released by Qwen on August 3, 2026, 63 days after Qwen3.7-Plus, as a 2.4 trillion parameter model with a 1 million token context window. Benchmark results span BullshitBench v2, PaperBench, OSWorld-Verified, BabyVision, SWE-Bench Pro, DeepSWE 1.1, and others. Notable scores include PaperBench at 93%, SW The model's competitive positioning places it 5th of 23 on SWE-Bench Pro behind Claude Fable 5.1 (81.2%) and 16th of 26 on DeepSWE 1.1 behind Muse Spark 1.3 (75.4%). It also tracks QwenSWEBench V2 results as part of Alibaba's in-house coding benchmark suite. Pricing data was fetched from openrouter.ai on September 20,

On July 19, 2026, Alibaba's Qwen team previewed Qwen3.8-Max-Preview during the World AI Conference in Shanghai, framing it as a 2.4-trillion-parameter multimodal flagship that the vendor claims ranks second only to "Fable 5" among systems it benchmarked. The preview went live immediately, though Alibaba did not publish For developers, the practical surface is hosted access rather than downloadable weights: the preview is reachable through Alibaba's Token Plan and Qoder/QoderWork tooling, with multimodal inputs (text, images, video, documents) and a 1-million-token context window. Open weights are promised "soon" without a date or lic

BenchLM.ai tracks Qwen3.8 Max with a capability score of 71.8/100 and a multimodal rank of 5, particularly strong for screenshots, documents, charts, and grounded multimodal workflows. The model has 60 published benchmark rows verified across categories including Agentic (15/15 verified), Coding (12/12 verified), Reaso The aggregator notes that no comparable first-party API token rate is published, with API pricing unavailable from the model's own documentation. Category rankings include Agentic at rank 9 of 154 (95th percentile), Coding at rank 20 of 153 (88th percentile), and Knowledge at rank 17 of 184 (91st percentile). Speed is

This integration note corroborates that Qwen3.8-Max-Preview was made available through Alibaba Token Plan's OpenAI-compatible endpoint under the model ID qwen3.8-max-preview, three days after Moonshot's Kimi K3, at WAIC in Shanghai. It reports a 2.4-trillion-parameter multimodal architecture with a 1-million-token cont The piece is the most developer-actionable source in the set: it documents that preview pricing is reportedly set at about 10% of standard Token Plan rates, names the concrete model identifier for API calls, and clarifies that "Qwen 3.8" is a family version rather than an 8-billion-parameter sibling, so existing on-dev