Qwen3.8 27B

Unifically's technical breakdown describes Qwen3.8 27B as a 27-billion-parameter dense causal language model with a vision encoder, built on the Qwen3.5 architectural foundation of 64 layers interleaving Gated DeltaNet linear-attention blocks with Gated Attention blocks, trained with multi-token prediction. Context is The article highlights a generation jump on LuminaBench's composite index, where Qwen's 27B moves from 42.2 (rank 67) for Qwen3.6 27B to 67.3 (rank 19) for Qwen3.8 27B, alongside GPT-5.4 at 67.4 and Qwen's hosted 3.7 Max at 67.6. It also notes caveats: nearly all launch numbers come from Qwen's own reporting, and Simon

DataNorth reports that Alibaba released Qwen3.8-27B on August 14, 2026 as an open-weights dense multimodal model accepting text, images, and video under the Apache 2.0 licence. The configuration file carries a full vision encoder and sets the language-model-only flag to false, clarifying that this checkpoint is a nativ On Alibaba's own evaluations, Qwen3.8-27B scored 61.7% on SWE-bench Pro against 53.4% for Claude Opus 4.6 Max, and Alibaba positions the model at local agent work where a 27B dense checkpoint fitting on a single high-end consumer or workstation GPU can run long-horizon tool calling that previously required hosted front

Alibaba's Qwen team unveiled Qwen3.8-27B on August 14, 2026 as an open-weight addition to the Qwen3.8 series, released under the Apache 2.0 license. The model is described as a native multimodal dense model that handles text, images, and video, with a 262K-token native context window extensible to one million tokens. W The blog also announces that Alibaba released weights for its flagship Qwen3.8-2.4T-A95B on Hugging Face and ModelScope, marking the first time a Qwen-Max-class model was open-released. Alibaba claims Qwen3.8-27B matches the performance of the Qwen3.7-Plus MoE model at roughly one-tenth its size, with strengths in codi

MLQ News confirms that Alibaba's Qwen team released Qwen3.8-27B as an open-weight dense multimodal model with a native 262,144-token context window, available through Qwen's Hugging Face and ModelScope releases under the Apache 2.0 license. According to the model card, the model supports image, text, and video inputs w The article explicitly flags that the release materials reviewed do not provide a complete like-for-like benchmark table covering Qwen3.8-27B, prior Qwen models, and competing open models under the same prompts, harnesses, and evaluation dates. As a result, claims of superiority are treated as Alibaba-attributed compan

Local AI Zone's technical analysis (published August 15, 2026) explicitly names Qwen3.8-27B and attributes the model to Alibaba's Tongyi Lab (Qwen Team) under an Apache 2.0 license, released August 14, 2026 at 15:00 UTC via Hugging Face. The article reports an exact parameter count of 27.78 billion (27,781,427,952) acr The analysis details a hybrid attention architecture using a 3:1 ratio of Gated DeltaNet (linear attention) to full attention layers — 48 Gated DeltaNet layers (75%) and 16 full attention layers (25%) — designed to reduce computational complexity from O(n²) to O(n) for 75% of layers. The piece frames Qwen3.8-27B as a c

Kingy.ai's launch-day review confirms Qwen3.8-27B was released at 15:00 UTC on August 14, 2026, with an exact checkpoint of 27.78 billion parameters accepting text, images, and video, shipping under Apache 2.0 with a native 262,144-token context window. The article documents large improvements over Qwen3.6-27B in agent The review offers critical caveats: every launch score comes from Qwen, several benchmarks are in-house or modified, and the SWE-bench Pro comparison imports Anthropic's Opus result rather than rerunning it under Qwen's setup. One million tokens require YaRN scaling rather than the native context, and Qwen warns that s

Latent.Space's AINews roundup covers the broader Qwen3.8 family announcement, noting Qwen 3.8 Max as a 2.4T-parameter model offered on API at $2 input/$6 output per million tokens with promised open-weights release for both Max and the 27B variant. The piece frames the releases in the context of renewed open-weights mo The roundup highlights Qwen3.8-Max capability claims including 10+ days of unattended autonomous coding building a self-evolving harness, a 125-hour autonomous research loop that invented a new data-selection method beating the referenced paper by +2.71 points, top 13% placement against 526 human teams in the WWW2025 M

Emergent reports that Alibaba Cloud's Qwen Team officially launched Qwen3.8-27B on August 14, 2026 as a 27-billion-parameter multimodal artificial intelligence model released under the permissive Apache 2.0 license. The model is positioned in the mid-range tier of large language models, designed to balance computationa The Apache 2.0 license grants unrestricted commercial use without royalty payments, permission to modify and distribute derivative works, patent protection for contributors and users, and no requirement to disclose source code modifications. Alibaba Cloud has emphasised cross-modal understanding in recent Qwen releases