Qwen3.8-Max

Shattered.io's Laura Bennett writes that Alibaba pushed Qwen3.8-Max onto Hugging Face and ModelScope in mid-August 2026, with the open-weight MoE packaged as Qwen3.8-2.4T-A95B landing on August 13 and the dense Qwen3.8-27B sibling following on August 14, while the hosted API debuted August 2 under Alibaba's Qwen.ai blo The architecture is described as a sparse mixture-of-experts design with 2.4 trillion total parameters and 95 billion active on any forward pass, positioned squarely at software engineering and multi-step agent tasks rather than general chat. Bennett reports a 86.6 score framing Qwen3.8-Max as "nearing Opus 5," and not

Rost Glukhov's Medium piece focuses on the Qwen3.8-27B open-weight sibling but centers it on the Qwen3.8-Max release, reporting that both checkpoints' weights have now been published on Hugging Face (Qwen/Qwen3.8-2.4T-A95B and Qwen/Qwen3.8-27B) and on ModelScope. The article describes Qwen3.8-Max as a 2.4-trillion-para The piece frames the open-weight rollout as the headline of the week for the local-AI community, quoting the official Qwen account's promise that open weights for Qwen3.8-Max and Qwen3.8-27B would ship "next week." It positions the larger model as a technological spectacle but emphasizes that the 27B checkpoint is the

Roboflow's Piotr Skalski reports on testing Qwen3.8-Max inside the Roboflow Playground against an upcoming VLM benchmark, finding that the model delivers the strongest object-detection results Roboflow has measured from a VLM and also performs well on counting and visual reasoning. The blog confirms Qwen3.8-Max is a 2. The piece details that Qwen3.8-Max handled satellite imagery, infrared imagery, documents, technical drawings, hand-drawn sketches, diagrams, crowded scenes, and images with small objects without task-specific training, while noting trade-offs around text extraction, latency, and deployment. Roboflow prompts the model

Alibaba's Qwen team released Qwen3.8-Max on August 3, 2026, making the new flagship model generally available through QwenCloud and Alibaba Cloud Model Studio, with open weights for Qwen3.8-Max and a sibling Qwen3.8-27B checkpoint scheduled to land on Hugging Face and ModelScope the following week. According to the Dat The release pairs Qwen3.8-Max with Qwen3.8-27B, a dense 27-billion-parameter checkpoint aimed at on-premise GPU hardware, marking Alibaba's first open-weight flagship in the Qwen-Max class. The article notes Alibaba first previewed the model in mid-July 2026 before the August 3 GA, which shipped alongside a full benchm

The Decoder reported on August 3, 2026 that Alibaba's Qwen team unveiled Qwen3.8-Max, a 2.4-trillion-parameter language model designed to complete complex tasks autonomously over multi-day horizons. The model activates 95 billion parameters per query and builds on the Qwen3.5 architecture, with internal benchmarks plac The coverage highlights three autonomous coding demonstrations: a 16-day build of the command-line tool oh-my-cli in which the model independently converted user requests into GitHub issues, wrote code, and iterated on tests; a chip design exercise; and a simulated e-commerce fiscal-year run. These case studies undersc

BenchLM's aggregator page, dated as of September 23, 2026, tracks Qwen3.8-Max across 60 published benchmark rows out of a broader catalog, scoring the model at 72/100 on its capability composite with a 1 million token context window and a measured decode speed of about 39 tokens per second with a 54.03-second first-tok BenchLM flags Qwen3.8-Max as particularly strong for screenshots, documents, charts, and grounded multimodal workflows, and notes that 60 published rows leave some tracked benchmark slots empty. The page explicitly states no comparable first-party hosted token rate is published, so neither Alibaba nor DigitalOcean API

Alibaba officially announced Qwen3.8-Max on August 3, 2026 as its largest and most capable flagship model to date. The model features 2.4 trillion total parameters with 95 billion activated per inference, built on a Sparse Mixture-of-Experts architecture combined with a hybrid attention mechanism atop the Qwen 3.5 foun The announcement positions Qwen3.8-Max as targeting frontier-level performance across autonomous coding, real-world applications, and long-horizon tasks, with a reported fourth-place ranking in the Frontend Code Arena. The sparse MoE design is framed as balancing massive scale with inference efficiency, delivering fron

DataCamp's blog, updated September 2, 2026, documents the Qwen3.8-Max-0902 post-training snapshot released that day, an upgrade focused on coding and long-horizon agent work that keeps the original architecture unchanged at 2.4 trillion parameters, 1 million token context, and the same $2 per million input and $6 per m Against Claude Opus 5, the 0902 snapshot leads on three coding benchmarks (MLS-Bench-Lite 50.1, SWE-Atlas QnA 66.3, QwenSWEBench V2 70.0) plus WorkArena Elo (1468) and both published multimodal evals (MMMU-Pro 82.7, ERQA 78.3), while Claude still leads TerminalBench 3.0 (42.7), DeepSWE 1.1 (73.6), and most agent coordi