Qwen3.8 Max

Aivancity's technical write-up confirms that Alibaba Cloud officially unveiled Qwen3.8-Max on August 3, 2026, describing it as the largest and most capable Qwen model to date, with 2,400 billion total parameters (95 billion active), a Mixture-of-Experts architecture, and a reported context window of one million tokens. For practitioners, the practical signal is that long-running, agent-style coding tasks and large-context workflows are the use cases Alibaba is targeting, with clear improvement on coding and certain professional benchmarks but no claim of overall superiority. The open-weight distribution under Qwen3.8-2.4T-A95B makes

Alibaba's Qwen3.8-Max debuted as a sparse Mixture-of-Experts model with 2.4 trillion total parameters and 95 billion active parameters, first going live through Alibaba's QwenCloud API on August 2, 2026. The open-weight variant, packaged as Qwen3.8-2.4T-A95B, was published on Hugging Face and ModelScope on August 13, w For developers evaluating coding agents and large-context workloads, the technical takeaway is the combination of a 1M-token context window, MoE efficiency that keeps active parameters manageable, and Apache 2.0 weights that make self-hosting feasible. The release lands amid a crowded September 2026 frontier cycle in w

On September 2, 2026, Alibaba released Qwen3.8-Max-0902, a post-training upgrade to Qwen3.8-Max that keeps the same 2.4 trillion parameter architecture and 1 million token context while focusing on coding and long-horizon agent performance. The new variant ranks first overall on Code Arena WebDev with 1,691 points, edg The Qwen3.8-Max-0902 update leads Claude Opus 5 on three coding benchmarks (MLS-Bench-Lite, SWE-Atlas QnA, QwenSWEBench V2) and wins on WorkArena and both published multimodal evaluations, though Claude retains leads on TerminalBench 3.0, DeepSWE 1.1, and several agent coordination benchmarks. Additional strong gains a

A third-party benchmark breakdown of Qwen3.8-Max published August 14, 2026 details the model's published scores: PaperBench 93.0, IFBench 82.8, Terminal Bench 2.1 at 86.6, MathVision 95.2, and OSWorld-Verified 86.1, alongside reported losses at HLE 43.6 (last among the four flagships compared) and SWE-bench Pro 67.7 (r The same piece provides context on reported pricing of $2 input and $6 output per million tokens (roughly a third of Claude Opus 5 and a quarter of GPT-5.6 Sol) and comparative framing against Kimi K3, Fable 5, and GPT-5.6 Sol, recommending the benchmark table be treated as a strong vendor claim rather than an independ

Roboflow confirmed that Alibaba's Qwen3.8-Max uses a 2.4-trillion-parameter MoE architecture that activates approximately 95 billion parameters per query, accepting text, images, and video as input and returning text. The model was available through the Alibaba Cloud API at the time of writing, with open weights scheduled for release on August 12, 2026, alongside a smaller dense Qwen3.8-27B model. Roboflow also tested Qwen3.8-Max in its Playground. In Roboflow's internal VLM benchmark, Qwen3.8-Max delivered the strongest object detection results among tested VLMs, performing well across satellite imagery, infrared images, documents, technical drawings, hand-drawn sketches, diagrams, crowded scenes, and small-object images. The model returned detections as normalized XYXY coordinates in a 0–1000 range, and accuracy was sensitive to coordinate order, scale, and response structure. These are single-vendor benchmark results rather than independent reproductions.

Roboflow's evaluation of Qwen3.8-Max for vision tasks confirms the model accepts text, images, and video as input and returns text, available through the Alibaba Cloud API and testable in Roboflow Playground. The model delivered the strongest object detection results in Roboflow's upcoming VLM benchmark, performing wel The vision workflow requires prompting the model to return a JSON list containing a label and four coordinates per detected object, which Roboflow then parses and converts into bounding boxes, with Qwen models working best on XYXY coordinates normalized to a range of 0 to 1000. Open weights were scheduled for release o

Alibaba officially announced Qwen3.8-Max on August 3, 2026, describing it as the most powerful model in its Qwen series to date. The model is a multimodal foundation built on a Qwen 3.5 base with a Sparse Mixture-of-Experts architecture and hybrid attention mechanism, totaling 2.4 trillion parameters with 95 billion ac The creator-attributed announcement confirms Qwen3.8-Max natively supports visual intelligence and is positioned as Alibaba's flagship, with availability also through QwenWork, Alibaba's all-in-one workplace AI agent platform. Ranked fourth on Fronted Code Arena, the model is designed for autonomous coding and long-hor

Qwen officially released Qwen3.8-Max, a 2.4-trillion-parameter Mixture-of-Experts model with 95 billion active parameters built on the architectural foundation of Qwen 3.5, marking the first time weights of a Qwen-Max-class model will be open-sourced (releasing the week following the announcement). The post, republishe The announcement positions Qwen3.8-Max against Opus 4.8, Fable 5, GPT-5.6 Sol, and Gemini 3.1 Pro with full benchmark tables, covering coding, RL infrastructure gains, and multimodal agent capabilities. The release emphasizes end-to-end, dependable delivery of complex tasks rather than single-turn response quality, and

Alibaba's Qwen3.8-Max is a flagship mixture-of-experts model with 2.4 trillion total parameters, designed for coding and long-horizon agentic workloads rather than local deployment. The explainer confirms multimodal capabilities, a context window of up to 1 million tokens, and selective expert activation per token to manage inference cost. The article positions Qwen3.8-Max as part of the shift from chatbots toward AI agents. The model is built to participate in extended workflows involving planning, tool use, result inspection, and follow-up actions rather than single-turn answers. Alibaba targets complex reasoning, software development, computer-use tasks, and long-running agentic pipelines with this release. The source is a third-party explainer and does not include independent benchmark reproduction, so figures should be read as Alibaba-reported.

A third-party benchmark aggregator (BenchLM.ai) published a Qwen3.8 Max profile dated October 2, 2026, listing a capability score of 72.1/100, a 1M-token context window, and a measured speed of 39 tok/s. The page ranks the model 16th of 212 tracked models and notes that no first-party hosted API token rate is published. It is presented as a tracked catalog entry rather than an official release. Per-category verified benchmarks show Multimodal at 88.4 (rank 5 of 49, 92nd percentile) and Instruction Following at 89.8 (rank 16 of 125, 88th percentile), with 24 multimodal rows and 15 agentic rows. The Aggregator reports 60 displayable benchmark rows with some tracked slots empty. Data is current to October 2, 2026.