Qwen3.8 27B

Qwen3.8-27B scores 52 on the Artificial Analysis Intelligence Index at maximum reasoning effort and 61.7 on SWE-bench Pro per Qwen's own evaluation, and leads its comparison set on agentic coding and computer-use benchmarks while trailing frontier models on Humanity's Last Exam and GPQA Diamond, according to a Qubrid A On vision-language workloads the same model card shows large gains over Qwen3.6-27B, with OSWorld-Verified moving from 63.9 to 84.3, WebArena-Verified from 48.8 to 64.8, AndroidWorld from 70.3 to 81.9, RecreationBench from 29.8 to 47.1, SWE-MM from 25.7 to 38.6, and Vision2Web from 45.0 to 62.9, while trailing Qwen3.7-

Regolo.ai's benchmark analysis reports that Qwen3.8-27B outperforms Claude Opus 4.6 Max on 16 of the 24 benchmarks in its official model card, with the largest wins in agentic coding (SWE-bench Pro: 61.7 vs 53.4), instruction following (IFBench: 79.5 vs 62.5), and computer use (OSWorld-Verified: 84.3 vs 72.7), while st Qwen3.8-27B is described as a dense, open-weight vision-language model with 27 billion parameters (28 billion including the vision encoder), a hybrid 48 Gated DeltaNet plus 16 full-attention layer architecture, a native 262,144-token context window extensible to 1 million with YaRN, image and video input, and the abili

A Medium article by Rost Glukhov titled "Qwen 3.8 27B Is Just Released - and It Could Be the Most Important Local AI Release of 2026" explicitly discusses the Qwen3.8-27B open-weights release and links the Hugging Face repositories (huggingface.co/Qwen/Qwen3.8-2.4T-A95B and huggingface.co/Qwen/Qwen3.8-27B). It frames t The article is opinion-framed commentary rather than primary technical reporting and contains no independent benchmark reproduction, limiting its evidence weight. It also mixes Qwen3.8-Max and Qwen3.8-27B coverage, and Medium's member-only truncation may obscure detail for non-subscribers. As community reception and ad

Five.reviews describes Qwen3.8-27B as a 27-billion-parameter dense multimodal model from Alibaba designed for open-weight deployment, released on August 14, 2026 under Apache 2.0 with native multimodal capabilities, 262K-token context (extended to 1M via YaRN), and reasoning and agentic modes. The model improves substa The article frames Qwen3.8-27B as Alibaba's response to what a well-trained 27-billion-parameter model can do with modern post-training techniques, built on the Qwen3.5 foundation with continued pretraining and fine-tuning on code, reasoning, and vision datasets, and optimized for software engineering, agent-based auto

Qwen3.8 27B is the compact model in Alibaba's Qwen3.8 open-weight generation, published on Hugging Face on August 14, 2026 under Apache 2.0 as a 27-billion-parameter dense model with native image and video input, 262K tokens of context (extensible to 1M via YaRN), and an FP8 checkpoint, according to Unifically. On Qwen Architecturally the article describes Qwen3.8 27B as a causal language model with a vision encoder built on the Qwen3.5 foundation: 64 layers interleaving Gated DeltaNet linear-attention blocks with Gated Attention blocks at a 3:1 ratio, trained with multi-token prediction, with 262,144 tokens of native context extensi

Sanj's deployment guide reports that Qwen3.8-27B landed on Hugging Face on 14 August 2026 and set a new bar for what a 27-billion-parameter dense model can do, leaving Qwen3.6-27B behind by double-digit margins on SWE-bench Pro, Terminal-Bench, and DeepSWE while staying small enough to run on hardware developers probab The same guide frames the upgrade over Qwen3.6 around agentic coding gains, gives concrete deployment commands for SGLang and vLLM, and offers INT4, FP8, and BF16 VRAM targets for running Qwen3.8-27B locally, with configuration, benchmark tables, and hardware sizing derived from the Qwen model card rather than independ

On August 14, 2026 at 15:00 UTC Alibaba's Tongyi Lab released Qwen3.8-27B on Hugging Face under Apache 2.0, a 27.78-billion-parameter dense multimodal language model that the Local AI Zone analysis says achieves competitive performance with models 10-15× its size while keeping deployment to a 24GB VRAM minimum. The tec The same analysis claims Qwen3.8-27B outperforms Meta's Muse Glimmer (30B) across all 8 direct comparison benchmarks and surpasses Claude Opus 4.6 on 15 of 19 overlapping tests, establishing what it calls a new efficiency frontier for local language models. Architecture details, hybrid attention mechanism, and key inno

Alibaba's Qwen team released Qwen3.8-27B at 15:00 UTC on August 14, 2026, and Kingy.ai's launch-day review calls it the most convincing candidate yet for the best dense, locally deployable multimodal model around 30 billion parameters. The checkpoint contains 27.78 billion parameters, accepts text, images, and video, s Kingy.ai flags that every launch score comes from Qwen, several benchmarks are in-house, corrected or modified, the SWE-bench Pro comparison imports Anthropic's Opus result rather than rerunning it under Qwen's setup, and Qwen warns that static YaRN scaling can hurt shorter prompts, while the Kingy.ai August 22 update