Qwen3.8-27B

Alibaba released open weights for both Qwen3.8-Max and Qwen3.8-27B on Hugging Face and ModelScope, with the 27-billion-parameter variant positioned as the practical local-deployment counterpart to the 2.4-trillion-parameter Max flagship. The official Qwen account announced that "Qwen3.8-Max will be released, and Qwen3. The piece frames Qwen3.8-27B as the "model that can leave the cloud," targeting developers who want to own, run, and customize their AI rather than relying on API access. It positions the 27B release as potentially the most important local AI release of 2026 for the open-weight community, contrasting it with the techno

The New Stack's mainstream tech-press article, published August 14, 2026, frames Qwen3.8-27B as delivering "Opus 4.6-level performance on your laptop," signaling strong developer interest in local-inference workflows for the newly released Alibaba model. The coverage ties into the broader narrative of frontier-capable As a press-style piece rather than a benchmark reproduction, the article provides useful launch-day positioning of Qwen3.8-27B within the AI tooling ecosystem, though the scraped excerpt is truncated by the site's cookie and consent UI, limiting directly visible technical depth from the page. Its value is as independen

Alibaba's Qwen team released Qwen3.8-27B as an open-weight model at 15:00 UTC on August 14, 2026. The checkpoint contains 27.78 billion parameters, accepts text, images, and video, ships under the Apache 2.0 license, and has a native 262,144-token context window. Qwen reports substantial gains over Qwen3.6-27B in agent The launch-day review notes important caveats: every score comes from Qwen, several benchmarks are in-house or modified, the SWE-bench Pro comparison imports Anthropic's Opus result rather than rerunning it under Qwen's setup, and the 1M-token context requires YaRN scaling rather than native context, with Qwen warning

On the LLM Stats aggregator, Qwen3.8-27B ranks 31 overall, with a Vision tier of "Good" (Top 10%) at 18 of 208 models, and average standing in Reasoning (32 of 363), Coding (41 of 267), Tool Calling (43 of 194), and Math (70 of 327). The model scores an LLM Stats composite of 45.1 at a blended price of $0.52 per millio The aggregator tracks multiple evaluation datasets for Qwen3.8-27B, providing a secondary index of official benchmark numbers attributed to the model's Hugging Face scorecard and official blog posts. It frames the model within a broader competitive landscape spanning Gemma 4 E4B, GPT OSS 120B, DeepSeek variants, and fl

The open-weights week Alibaba promised for its Qwen3.8 generation split on August 13, 2026. The Max-class flagship's weights, Qwen3.8-2.4T-A95B, a 2.4-trillion-parameter sparse mixture-of-experts model, shipped on Hugging Face and ModelScope as the first Max-level Qwen release ever made downloadable. However, Qwen3.8-2 Third-party trackers described the Qwen3.8-27B release as delayed with no new date given, raising questions about why the Max's weights shipped first and what happened to the 27B. The page tracks the open-weight release status as a developing story, noting the gap between Alibaba's commitment and actual delivery for th

A community thread on the NVIDIA Developer Forums (posted Aug 8, 2026) explicitly announces that Qwen3.8-27B is coming the following week and notes that the full Qwen 3.8 family will be released as open weights, pointing readers to the official Qwen blog at qwen.ai/blog?id=qwen3.8 for the source announcement. The post As an informal forum item rather than an official Qwen or evroc source, this thread should be treated as community corroboration of the release context for Qwen3.8-27B rather than primary documentation. It does not provide architectural or benchmark details, but its explicit naming of Qwen3.8-27B alongside the open-wei

Local AI Zone's paper-style analysis confirms the August 14, 2026, 15:00 UTC release of Qwen3.8-27B by Alibaba's Tongyi Lab as a 27.78-billion-parameter (27,781,427,952 parameters) dense multimodal model under Apache 2.0, claiming it outperforms Meta's Muse Glimmer (30B) across all 8 direct comparison benchmarks and su Key architectural innovations documented include a hybrid 3:1 attention ratio with 48 Gated DeltaNet (linear attention) layers and 16 Gated Full Attention layers, reducing computational complexity from O(n²) to O(n) for 75% of layers, plus a built-in multi-token prediction (MTP) head for speculative decoding, an attent