Qwen3.8 27B

A DEV Community deep-dive from 19 August 2026 documents Qwen3.8-27B as a native vision-language model — not a text model with a bolted-on multimodal adapter — that handles text, images, and hour-long video, with reasoning enabled by default and toggleable via a reasoning_effort parameter supporting xhigh, medium, and l The piece provides the most detailed architecture breakdown supplied across the candidates: a hybrid linear/full-attention stack with 27B parameters (about 28B including padding), hidden dimension 5,120, 64 layers in a 16×(3×Gated DeltaNet → FFN → 1×Gated Attention → FFN) pattern, 48 Gated DeltaNet value heads and 16 Q

Rost Glukhov's Medium post reports that Alibaba's Qwen team released Qwen3.8-27B open weights on Hugging Face (huggingface.co/Qwen/Qwen3.8-27B) and ModelScope, positioning it as a locally deployable alternative to the much larger Qwen3.8-Max flagship. The article notes the announcement originated from the official Qwen The piece frames Qwen3.8-27B as potentially the most important local AI release of 2026 for developers who want to own, run, and customize their AI rather than rely on cloud APIs. It highlights the shift from closed frontier models toward open-weight availability for the Qwen3.8 family, giving the developer community d

Alibaba released Qwen3.8-27B on 14 August 2026 as an open-weights dense model under the Apache 2.0 license, with the checkpoint published on Hugging Face. According to the supplied DataNorth report, the model natively accepts text, images, and video and returns text, with a 262,144-token native context window that Alib The same report cites Alibaba's own evaluations scoring Qwen3.8-27B at 61.7% on SWE-bench Pro, compared with 53.4% for Claude Opus 4.6 Max, and describes the model as targeting local agent workloads where a 27B dense checkpoint that fits on a single high-end consumer or workstation GPU can handle long-horizon tool call

DEV Community's guide describes Qwen3.8-27B as a 27-billion-parameter dense language model with integrated vision capabilities, built by Qwen on the architectural foundation of Qwen3.5. The model features 64 layers, a hidden dimension of 5120, and a hybrid attention architecture combining Gated DeltaNet (48 linear atte The guide highlights Qwen3.8-27B's benchmark performance, including 61.7% on SWE-bench Pro, 73.0% on Terminal Bench 2.1 (Terminus), 42.2% on DeepSWE 1.1, 42.3% on NL2Repo-Bench, 84.3% on OSWorld-Verified (computer use), and 64.8% on WebArena-Verified (browser use), positioning it as well-suited for software engineering

The New Stack covers the release of Alibaba's Qwen3.8-27B, framing it as a model promising Opus 4.6-level performance on local laptop hardware. The article focuses on the model's capability for local inference deployment, emphasizing that Alibaba is making frontier-class AI weights available for self-hosted scenarios r The piece situates Qwen3.8-27B within the broader competitive landscape of open-weight versus closed-API frontier models, noting that the ability to run such a capable model locally represents a shift in how developers and organizations can access advanced AI capabilities. The coverage emphasizes implications for local

According to kingy.ai's launch-day review, Alibaba's Qwen team released Qwen3.8-27B on August 14, 2026, at 15:00 UTC, describing it as a dense, locally deployable multimodal model around 30 billion parameters. The checkpoint is reported to contain 27.78 billion parameters, accept text, images and video, ship under Apac The page reports official model-card benchmark gains including Terminal-Bench 2.1 rising from 63.4 to 73.0, DeepSWE 1.1 from 13.3 to 42.2, OSWorld-Verified from 63.9 to 84.3, and SWE-MM from 25.7 to 38.6, but the reviewer flags that all launch scores come from Qwen, that several benchmarks are in-house, and that one mi

Local AI Zone's comprehensive technical analysis documents that Alibaba's Tongyi Lab released Qwen3.8-27B on August 14, 2026, as a 27.8-billion-parameter dense multimodal language model with Apache 2.0 licensing. The model employs a hybrid attention architecture with a 3:1 ratio of Gated DeltaNet (linear attention, 48 The analysis reports Qwen3.8-27B achieves competitive performance with models 10-15× its size while maintaining practical deployment requirements (24GB VRAM minimum), outperforming Meta's Muse Glimmer (30B) across all 8 direct comparison benchmarks and surpassing Claude Opus 4.6 on 15 of 19 overlapping tests. The autho