Benchmarks · AUGUST 18, 2026
Qwen3.8-27B lands at 3M downloads and matches GPT-5.6 Luna on Artificial Analysis
Alibaba's 27.78B-parameter dense multimodal model shipped August 14 under Apache 2.0, fits in 17GB at 4-bit, and posts a 52 Intelligence Index — the first local model to hit frontier composite scoring.
Alibaba's Qwen team shipped Qwen3.8-27B on Hugging Face at 15:00 UTC on August 14, and by Sunday the checkpoint had crossed 3 million downloads under Apache 2.0. That number, per Cybernews, is remarkable less for its size than its composition: Unsloth's quantizations pulled roughly 2 million of it, LMStudio's community builds another 500,000. The people grabbing this model aren't cloud engineers. They're running it on their own GPUs.
Which is the actual story. A 27.78B-parameter dense multimodal model with a 262,144-token context posts a 52 on Artificial Analysis's Intelligence Index, the same composite score OpenAI's GPT-5.6 Luna hits at maximum reasoning. Artificial Analysis builds that number from nine evals spanning coding, science, reasoning, and professional tasks. No open-weights model under 30B has landed inside that band before.
The agentic side tells a similar story. Qwen3.8-27B scores 51 on the Agentic Index, above Claude Opus 4.8 at maximum reasoning. On Alibaba's own runs against an amended SWE-bench Pro (the team corrected problematic tasks and re-scored all baselines, per DataNorth), it posts 61.7% to Claude Opus 4.6 Max's 53.4%. LiveCodeBench v6: 90.3% vs 88.8%. IFBench: 79.5% vs 62.5%. It trails on GPQA Diamond (89.2% vs 91.3%) and Humanity's Last Exam (30.8% vs 40.0%). Alibaba also amended ground-truth annotations on MathVision and CharXiv.
Vendor-run harnesses on vendor-amended test sets. That's the asterisk. Kingy.ai notes no independent reproduction as of launch day.
The generational jump against Qwen3.6-27B is the more legible number, because it's apples-to-apples: DeepSWE 1.1 from 13.3% to 42.2%, Terminal-Bench 2.1 from 63.4 to 73.0, OSWorld-Verified from 63.9 to 84.3, SWE-MM from 25.7 to 38.6. "The launch evidence shows a consistent upgrade over Qwen3.6-27B rather than one isolated benchmark spike."
There's a token-cost footnote worth pricing in. Qwen3.8-27B ran the Intelligence Index at its default "xhigh" reasoning effort and burned 160 million output tokens across the suite, against a 43 million median for comparable open-weight models. The 52 is real; it's also expensive to reproduce.
Hardware is where the release becomes structurally interesting. Full BF16 wants roughly 56GB of GPU memory. FP8 halves that to 28GB. 4-bit quantization drops it to about 17GB, RTX 3090 territory, which is why Unsloth's builds moved the download counter faster than the official checkpoint. The frontier composite band is now something a hobbyist can run on a five-year-old consumer card, offline, under a permissive license. The competitive moat OpenAI and Anthropic have been pricing against isn't the score anymore. It's what you're allowed to do with the weights once you have them.