Qwen3.8-27B

Qubrid's benchmark compilation shows Qwen3.8-27B scoring 52 on the Artificial Analysis Intelligence Index at maximum reasoning effort and 61.7 on SWE-bench Pro per Qwen's own evaluation, with the 27B leading its comparison set on agentic coding and computer-use benchmarks while trailing frontier models on Humanity's La On vision-language tasks, the model card figures show OSWorld-Verified at 84.3 (up from 63.9 on the prior 27B), WebArena-Verified at 64.8, AndroidWorld at 81.9, RecreationBench at 47.1, SWE-MM at 38.6, and Vision2Web at 62.9. The article notes that the published table places Qwen3.8-27B ahead of the listed Claude Opus

Rost Glukhov's Medium piece (publish date August 5, 2026 inside the article; surfaced August 18, 2026) reiterates the official Qwen account's announcement that the open weights of both Qwen3.8-Max and Qwen3.8-27B would be released, and confirms that the weights subsequently appeared on Hugging Face at Qwen/Qwen3.8-27B The piece is a personal Medium blog rather than an official source and provides limited new technical depth beyond confirming the open-weights distribution channels and emphasizing the local-AI positioning of the 27B checkpoint relative to the Max variant.

A Hacker News discussion highlights Qwen3.8-27B scoring 52 on Artificial Analysis, beating all medium models in the 40B–150B range and tying DeepSeek V4 Flash 0731, which ranks 5th in the large-model category (150B+). As a reference point, Qwen3.6 27B scored 38, which was previously the highest in the small-model categ Technical discussion in the thread notes Qwen3.8-27B's token usage is roughly 2.3× GPT Luna Max and almost 2× Kimi K3, raising questions about whether hosted providers can compensate through high tokens-per-second. Community members observe that Qwen models can be slower than comparable Gemma4 variants but are well-sui

Alibaba's Qwen team released Qwen3.8-27B on August 14, 2026, and the open-weight model quickly drew major adoption attention, racking up over 3 million Hugging Face downloads within its first weekend and becoming the number-one trending model on the platform at the time of reporting. The release shipped under an Apache Community reception emphasized the model's unusually strong capability-to-size ratio, with independent developer Simon Willison quoted saying it was "a miracle" that a 17GB file could deliver this performance on home machines. Alibaba positions Qwen3.8-27B as delivering near-frontier results, and the report explicitly

Local AI Zone's technical analysis describes Qwen3.8-27B as a 27.8-billion-parameter dense multimodal model released August 14, 2026 by Alibaba's Tongyi Lab under Apache 2.0, with a 24GB VRAM minimum for deployment. The architecture uses 64 transformer blocks with a hidden dimension of 5,120, 24 query heads and 4 KV he The model employs a hybrid attention mechanism with a 3:1 ratio, layering 48 Gated DeltaNet linear-attention layers against 16 full-attention layers, reducing computational complexity from O(n²) to O(n) for 75% of layers while preserving quality on long-context tasks. Native context is 262,144 tokens, with extended 1M-

Kingy.ai's launch-day technical review confirms Alibaba's Qwen team released Qwen3.8-27B at 15:00 UTC on August 14, 2026, and describes it as a dense 27.78-billion-parameter multimodal checkpoint accepting text, images, and video, shipped under Apache 2.0 with a native 262,144-token context window. Extending to one mil The article lists Qwen-reported benchmark gains over Qwen3.6-27B, 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, while explicitly flagging that all launch-day scores come from Qwen and that several comparisons (such

The LLM Stats aggregator page for Qwen3.8-27B (dated August 14, 2026) places the model at an overall rank of 32 with a composite LLM Stats Score of 45.1 and a blended price of $0.52 per million tokens, with capability tier rankings including Vision in the top 10% (18 of 208), Reasoning at 32 of 362, Coding at 40 of 266 The page lists individual benchmark scores sourced from the model's Hugging Face scorecard: MathVision 0.95 (rank 4), OmniDocBench 1.5 with an Overall Edit Distance score of 0.91 (rank 4, lower is better), and LiveCodeBench v6 at 0.90 (rank 5). Readers should note that the blended price point shown reflects serving eco

An NVIDIA DGX Spark / GB10 user-forum thread announced that Qwen3.8-27B would arrive the following week with full open-weights release, pointing to Alibaba's official Qwen blog post on the 3.8 family. The announcement was cross-posted from a LinkedIn post by Qwen, and the thread links the official qwen.ai blog. This ca Related threads visible in the forum index discuss deployment experiments specific to Qwen3.8-27B on DGX Spark hardware, including NVFP4 quantization with 1M context via vLLM+MTP, single-DGX-Spark optimizations for MixedInt4-AutoRound, and Flash-Next variants at 43 tok/sec for coding. These companion threads indicate a