Qwen3.8 27B
A 27-billion-parameter open-weight Qwen model under the Apache 2.0 license, with about 7.8 million downloads on Hugging Face. It is far below frontier models on hard tests, but its size and license make it a common choice for self-hosting.
Alibaba (Qwen)Language modelReleased August 2026Data checked October 18, 2026
This model is not in the current people's rankings: it is either new and still collecting votes, or ranked below the top 12. See the AI leaderboard for the models people rate highest.
On paper
Fixed-test rankings from the benchmarks tab: Speed #21, Reasoning #23.
Strengths and weaknesses
- Apache 2.0 license, so it can be used commercially and modified freely
- 42.4% on ARC-AGI-2 at its highest effort setting
- Cheap when hosted by Alibaba Cloud ($0.50 / $3 per million tokens)
- Low Artificial Analysis Intelligence Index (33.7) and weak at agentic terminal tasks (5.6% on Terminal-Bench 4.0)
- Near the bottom of this list on LMArena creative writing
Best for
- Self-hosting with a permissive license
- Private, offline use
- Low-cost experiments
Every score we have
Each bar compares this model with the best result any model on this page has on the same test.
- LMArena text arena (human votes)1,437.8Arena score 1437.8 (95% CI 1432.6 to 1442.9); 23,294 votes; entry "qwen3.8-27b"; position 100 of 416 on the full board
- LMArena creative writing1,366.4Arena score 1366.4 (95% CI 1356.8 to 1376.0); 4,668 votes; entry "qwen3.8-27b"; position 160 of 414 on the full board
- LMArena WebDev (coding)1,593.2Arena score 1593.2 (95% CI 1586.6 to 1599.8); 15,940 votes; entry "Qwen3.8-27B"; position 29 of 144 on the full board
- LMArena vision1,244.5Arena score 1244.5 (95% CI 1236.0 to 1252.9); 7,013 votes; entry "qwen3.8-27b"; position 61 of 165 on the full board
- LMArena Agent Arena-0.034IPS score -0.0343 (95% CI -0.0423 to -0.0264); 60,515 sessions; entry "Qwen3.8-27B"; position 37 of 52 on the full board
- Artificial Analysis Intelligence Index33.733.7 on the Artificial Analysis Intelligence Index v4.3.2 (AA shows 34), setting "Qwen3.8 27B (xhigh)"
- Humanity's Last Exam (Artificial Analysis)33.9%33.9% on Humanity's Last Exam (AA, 2,158 text-only questions), setting "Qwen3.8 27B (xhigh)"
- GPQA Diamond, graduate-level science90.5%90.5% on GPQA Diamond (AA; no longer part of the AA index since v4.2), setting "Qwen3.8 27B (xhigh)"
- Terminal-Bench 4.0 (Artificial Analysis run)5.6%5.6% on Terminal-Bench 4.0 as run by AA (mini-swe-agent harness), setting "Qwen3.8 27B (xhigh)"
- Long-context reasoning (AA-LCR)82.0%82.0% on AA-LCR (long-context reasoning), setting "Qwen3.8 27B (xhigh)"
- MMMU-Pro, image understanding76.3%76.3% on MMMU-Pro (image understanding, AA), setting "Qwen3.8 27B (xhigh)"
- AA-Omniscience, accuracy minus hallucination-10-10.0 on the AA-Omniscience Index (knowledge accuracy minus hallucination, range -100 to 100), setting "Qwen3.8 27B (xhigh)"
- Output speed, tokens per second41 t/s41 output tokens per second (AA median), row "Qwen3.8 27B (xhigh)"
- Cost per Artificial Analysis test task$1.01$1.01 to run one AA Intelligence Index task, row "Qwen3.8 27B (xhigh)"
- ARC-AGI-2 puzzles42.4%42.36% on ARC-AGI-2 (semi-private set), best listed setting "Qwen3.8-27B (XHigh)", $0.45 per task
Compare with
- Claude Opus 5.5Anthropic's flagship for long-running agentic coding and knowledge work, and the model Anthropic suggests starting with for most jobs.
- Claude Sonnet 5.5Anthropic's mid-priced model, which Anthropic describes as its best mix of speed and intelligence.
- Claude Fable 5.1Anthropic's top tier for the hardest reasoning and long-horizon agent work, which Anthropic recommends when Opus 5.5 still falls short.
- Claude Haiku 5.5Anthropic's smallest and fastest current model, built for high-volume, latency-sensitive jobs like classification, extraction and routing.