qwen3.8-max
Meant to excel at
-
Alibaba pitches long autonomous coding runs; an internal test ran a 16-day project producing an open-source agent framework.
Vendor claim, relayed alibabagroup.com -
Natively multimodal: turns screenshots into frontend code, floor plans into 3D interiors, and handles long documents, video and livestreams.
Vendor claim, relayed alibabagroup.com -
Targeted at legal document review, financial research and similar professional work (vendor claim, no figures).
Vendor claim, relayed alibabagroup.com -
Artificial Analysis index 58 (fifth overall), second on Vision Arena, top score on the tau3-Banking agent test.
Independent evaluator deeplearning.ai -
Front-end generation is its strongest hands-on area: responsive layouts, animation and 3D maps.
Independent evaluator mindstudio.ai -
Breaks prompts into ordered task lists and adds edge cases and security checks unprompted, similar to Opus.
Independent evaluator getneotiler.com -
Arena WebDev 1672 to 1674 (preliminary), rank 11 to 12, ahead of Kimi K3 and GLM-5.3.
Independent evaluator arena.ai
Watch out for
-
Slow and token-heavy: cost per benchmark task about $1.13, double Qwen3.7-Max, despite low per-token prices.
Independent evaluator deeplearning.ai -
Slowest model one reviewer had tested; tasks took 30 to 80 minutes in the preview.
Practitioner report thomas-wiegold.com -
Sparse documentation (no technical report or model card); open Max weights are text-only with reduced context; custom license terms.
Independent evaluator deeplearning.ai -
Can resurface finished to-do items from earlier prompts and occasionally refuses benign requests like a simple curl.
Practitioner report getneotiler.com
Benchmarks
Coding percentile 0.73
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| code arena webdev | 1688 | 0.86 | Independent evaluator cryptobriefing.com | 2026-09-03 |
| lmarena code arena webdev | 1674 | 0.61 | Independent evaluator arena.ai | 2026-10-19 |
Agentic coding percentile 0.55
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| swe-bench pro | 67.7 | 0.75 | Vendor's own site huggingface.co | 2026-08-13 |
| terminal-bench 2.1 | 86.6 | 0.46 | Vendor's own site huggingface.co | 2026-08-13 |
| terminal-bench 2.1 (aggregator, anotherwrapper.com) | 86.6 | 0.46 | Independent evaluator anotherwrapper.com | 2026-10-19 |
Overall percentile 0.42
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| artificial analysis intelligence index | 45 | 0.35 | Independent evaluator artificialanalysis.ai | 2026-10-19 |
| vals index | 66.1 | 0.50 | Independent evaluator latent.space | 2026-08-03 |
Reasoning percentile 0.36
| Benchmark | Score | Percentile | Source | Date |
|---|---|---|---|---|
| gpqa diamond | 92.6 | 0.33 | Vendor's own site huggingface.co | 2026-08-13 |
| humanity's last exam | 43.6 | 0.39 | Vendor's own site huggingface.co | 2026-08-13 |
Percentile is the share of tracked models this model beats on that benchmark, averaged across the group. Only benchmarks at least four tracked models share are scored.
Practitioner reports
- +/-
Wrote the longest files in the HTML animation test, about 540 lines on average.
Compared with kimi-k3 | 2026-10-04
Practitioner report empiriolabs.ai - +/-
Planned better than Gemini 3.1 Pro and added unprompted security checks, but pulled finished tasks into new work and randomly refused a simple curl.
Compared with gemini-3.1-pro | 2026-09-16
Practitioner report getneotiler.com - +/-
A coding-agent review notes benchmark gains (DeepSWE, FrontierSWE) were not independently reproduced as of Aug 24, preview model IDs silently re-route to production, and the 131K output cap forces chunking of large refactors.
Task: refactor | 2026-08-24
Practitioner report verdent.ai - +/-
Reviews relayed by a secondary source say it excels at responsive layouts and animated SVG but struggles with polished high-detail 3D and has slow output.
2026-08-20
Practitioner report ownlife.dev - -
The open-weight checkpoint disappointed the community: it is text-only, forces thinking mode and has 262K native context, so it is not equivalent to the hosted service that the benchmarks measured.
Task: chore | 2026-08-12
Practitioner report explainx.ai - +/-
A creator's $32 test of hosted Qwen3.8 Max had one standout (a 2,000+ line C++ simulation) and a polished wedding site, but several expensive misses (failed printable-engine design, weak game, generic image-to-site); concludes it is a candidate generator needing strong verification, not a production migration.
Task: feature | 2026-08-03 older than 90 days
Practitioner report ai.joaoqueiros.com - +/-
On release, one commentator said open models are close to the top closed models on agentic coding but not yet there; others questioned whether the price gap is real given how hard 2.4T models are to serve.
Task: agentic-long-running | Compared with claude-fable-5 | 2026-08-03 older than 90 days
Practitioner report latent.space - +/-
Produced a very thorough coffee-roaster site and one-shot a Go poker simulation, but was the slowest model he had tested and the preview endpoint may change.
Compared with grok-4.5 | 2026-07-20 older than 90 days
Practitioner report thomas-wiegold.com - +/-
Front-end work was strongest (restrained site, animated Pokemon encyclopedia, 3D map) but an ISS globe rendered badly and a crowd-text animation was faked with static text.
undated
Practitioner report mindstudio.ai