Qwen3.8 27B hits 52 on Artificial Analysis, beating Llama 2 13B

Qwen3.8 27B posted a 52‑point score on the Artificial Analysis benchmark, the top result among sub‑35‑billion‑parameter models released this year [Artificial Analysis][Hacker News].

What happened

Alibaba Cloud released the model in March 2026 and ran it on the benchmark’s 12‑task suite, which includes MMLU, GSM‑8K and HumanEval. The test set used 5 k prompts per task, and the score reflects a weighted average of accuracy and token‑efficiency metrics. The 52 rating beats Llama 2 13B (48) and Gemini 1.5 Flash 27B (49) and trails only Llama 3 70B (58) and Claude 3 70B (57) on the same leaderboard [Artificial Analysis].

Why it matters

At 27 B parameters, Qwen3.8 delivers a 4‑point gain over the 13 B Llama 2 baseline, showing better compute efficiency than many Western models that need 2–3× the parameters for comparable scores. Its placement raises the count of Chinese‑origin models in Artificial Analysis’s top‑20 from three to five, signaling growing acceptance of non‑Western training pipelines. Because cloud providers charge inference by GPU hour, a 27 B model that matches a 34 B model’s accuracy can cut hardware spend by roughly 20 % according to the benchmark’s token‑efficiency data [Artificial Analysis].

Editor’s take

The 52‑point result proves that Chinese‑scale training can match top‑tier Western offerings without inflating parameter counts, giving engineers a cost‑effective alternative and forcing the market to reassess the dominance of OpenAI‑ and Anthropic‑backed models.