Qwen3.8 27B

Ofox published a first-party deployment guide (Aug 16, 2026) for running Qwen3.8-27B locally with GGUF quants. The guide states the model was released by Qwen under Apache 2.0 on Hugging Face as a dense 27B vision-language model with a 262,144-token native context, and that the 17 GB figure circulating in the community The same guide reports Qwen3.8-27B total memory usage spans 11-19 GB across quants (2-bit 11-13 GB, rising to roughly 19 GB at 4-bit) and notes that the full 256K context will not fit on consumer hardware even with a 24 GB card, so KV-budget management is required. It explicitly states Qwen3.8-Max has no open weights a

Ofox's blog post on Qwen3.8-Max (Aug 3, 2026) confirms Qwen3.8-27B as part of the same release wave, stating Alibaba announced the 27B alongside the Max flagship as an open-weights release. It documents that on Ofox the Max model is served as bailian/qwen3.8-max over both OpenAI and Anthropic protocols, indicating Ofox The post positions Qwen3.8-Max as a 2.4-trillion-parameter model with 95B active parameters per token, built on the Qwen 3.5 architecture, with headline capabilities around long-horizon autonomous work. It states the Max open-weights release was "announced for next week" without a published license or date, meaning the

Qubrid's August 31, 2026 analysis of Qwen3.8-27B reports the model scores 52 on the Artificial Analysis Intelligence Index at maximum reasoning effort and 61.7 on SWE-bench Pro per Qwen's own evaluation, while leading its comparison set on agentic coding and computer-use benchmarks. The article reproduces Qwen's offici On the vision-language side, Qubrid's table records Qwen3.8-27B achieving OSWorld-Verified 84.3, WebArena-Verified 64.8, AndroidWorld 81.9, RecreationBench 47.1, SWE-MM 38.6, and Vision2Web 62.9, clearly labeling which numbers come from Qwen's internal evaluation versus independent measurement. The article's central fr

A Medium piece by Rost Glukhov dated August 2026 frames Qwen3.8-27B as the most important local-AI release of the period, contrasting the 27-billion-parameter open-weight model with Alibaba's 2.4-trillion-parameter Qwen3.8-Max flagship. The article quotes the official Qwen account announcing that "Next week, the open w Beyond the release announcement, the article positions Qwen3.8-27B within the community context of running and customizing AI locally rather than depending on cloud APIs. It notes that the 2.4T Max model is described in official Alibaba documentation as offering substantial improvements over Qwen3.7-Max in coding, prof

The New Stack's coverage from August 14, 2026 reports on Alibaba's release of Qwen3.8-27B, highlighting its capability to deliver performance comparable to Anthropic's Claude Opus 4.6 while running on consumer laptop hardware. The article frames the release as a milestone for local inference, noting the model's open-we Coverage emphasizes Qwen3.8-27B's multimodal capabilities and its 262K-token native context window, making it suitable for local deployment scenarios that previously required cloud API access. The piece positions the model as advancing the practical frontier of locally deployable AI by combining frontier-tier benchmark

Alibaba's Qwen team released Qwen3.8-27B at 15:00 UTC on August 14, 2026, as a 27.78-billion-parameter multimodal checkpoint under Apache 2.0 with a native 262,144-token context window that accepts text, images, and video. The launch-day review by Kingy.ai documents Qwen-reported benchmark gains over Qwen3.6-27B, inclu Kingy's verdict positions Qwen3.8-27B as a leading dense, locally deployable multimodal model around 30 billion parameters, calling it "the most convincing candidate yet" for that profile, but explicitly cautions it is not an honest one-for-one replacement for stronger frontier systems. The article highlights KV-cache

InnFactory's model directory page, updated September 3, 2026, catalogs the Qwen 3.8 family and identifies Qwen3.8-27B as the Apache 2.0 self-hosting pick with a 262k native context window (extensible to 1M for sibling variants like Qwen3.8-Max and Qwen3.8-Flash), while Qwen3.8-Flash-Next (125B/6B active) uses the qwen- The page also contextualizes Qwen3.8-27B alongside its siblings in the Qwen 3.8 lineup and addresses EU hosting considerations including GDPR availability, making it useful for European developers assessing sovereign self-hosting options. It reports vendor benchmarks such as SWE-bench Pro 61.7 and LiveCodeBench v6 figu

A comprehensive technical analysis of Qwen3.8-27B documents the model's release on August 14, 2026 by Alibaba's Tongyi Lab under Apache 2.0 license. The model is a 27.78-billion-parameter dense multimodal language model built on the Qwen3.5 foundation, accepting text, images, and video input. Key architectural details The analysis highlights Qwen3.8-27B's native 262,144-token context window extendable to 1M tokens via YaRN positional scaling, along with built-in multi-token prediction for speculative decoding support. According to the report, the model achieves competitive performance against models 10-15× its size while maintaining