Qwen3.8 Max: BenchLM's Source-Verifiable Benchmark Ledger

One-sentence takeaway

BenchLM normalizes Qwen3.8 Max's 52 source-displayable benchmarks to 79.91/100, ranking it sixth among 218 models, but its category ledger also shows strong performance in Reasoning, Agentic, Multimodal, and instruction following while AutomationBench, OSWorld 2.0, HLE, and other projects remain clear gaps; “sixth” cannot replace selecting tasks by component score.

Use cases

• Suitable tasks: Filtering models by capability category, checking the source of each score, and comparing Qwen with other models in the same visible evidence ledger.

• Unsuitable tasks: Treating the BenchLM overall score as a new experiment, or treating provider-exact official scores as results independently rerun by BenchLM.

• Applicable model versions: Qwen3.8 Max (the page marks its release as 2026-08-03); data freeze/refresh point: 2026-08-17.

• Applicable clients, Agents, or APIs: The BenchLM/BenchAlign database; actual deployment still requires returning to the official Qwen endpoint and one's own harness.

• Recommended reasoning tier and parameters: The BenchLM page does not disclose uniform Qwen call parameters; original configurations for each benchmark should be taken from its Evidence link and Qwen's official footnotes.

Test environment and workflow steps

1. BenchLM's methodology page says that public tables display exact-source rows only by default; generated or inferred benchmark values do not enter evidence, and each model-page row retains score, weight, cohort, and evidence state.

2. Overall Agentic/Coding uses the frozen BenchAlign v5.2 artifact; the other categories remain evidence views until each is validated.

3. Open the model page and first record the overall score, ranking, number of source rows, and category scores, then expand categories such as Coding, Agentic, Reasoning, Knowledge, Multimodal, and Instruction Following.

4. Open each row's original Evidence link to confirm the official Qwen release table, evaluation footnotes, harness, and date; the BenchLM aggregate ranking must not be treated as raw data.

Results data

Category overview

• Overall: 79.91/100, #6/218; 52 source-displayable rows.

• AgenticRank: 79.2, #2/131, 99th percentile, 14 benchmarks, 22% weight.

• CodingRank: 66.6, #11/135, 93rd percentile, 7 benchmarks, 20% weight.

• ReasoningRank: 94.3, #1/2, 100th percentile, 2 benchmarks, 17% weight.

• Knowledge: 66.5, unranked, 4 benchmarks, evidence verified.

• Multimodal: 88.1, #2/33, 97th percentile, 24 benchmarks, 12% weight.

• Instruction Following: 97.9, #1/40, 1 benchmark, 5% weight.

• Math and Multilingual: marked as unmeasured on the page; they cannot be inferred from the overall score.

Component readings

• Coding: SWE-bench Pro 67.7% (12.6 percentage points below the page's best verified row); Terminal-Bench 2.1 at 86.6%; PaperBench 93.0% is the current best verified row.

• Agentic: OSWorld-Verified 86.1%, CoWorkBench 74.8%, WideResearch 81.9%, AndroidWorld 85.3%, and MobileWorld 77.8% are all listed on the page as Qwen's current best verified rows; AutomationBench at 27.3% is 20.9 percentage points below GLM-5.3's 48.2%; OSWorld 2.0 at 19.4% is 51.2 percentage points below Claude Opus 5's 70.6%.

• Reasoning/Knowledge: LongBench v2 66.3%; MRCRv2 92.9%, 0.7 percentage points below the page's best; HLE 43.6%, 21.1 percentage points below Claude Opus 5's 64.7%; GPQA 92.6%, 2.9 percentage points below the page's best.

• Multimodal: MMMU-Pro 82.3%; OmniDocBench 1.5 at 92.1%; Video-MME (with subtitles) 90.4%; LVBench 81.8%, 3.6 percentage points below Gemini 3.7 Flash's 85.4%.

Conclusions

BenchLM's value lies in breaking “high score” into clickable source rows: Qwen3.8 Max's strengths are concentrated in long context, tool/desktop agents, multimodal documents and video, and instruction following, while automation-boundary tasks, OSWorld 2.0, and expert-knowledge exams must not be hidden by the overall score. For model selection, use the component corresponding to the business task, and save the Qwen official score's harness, temperature, repetition count, and footnotes together.

Limitations

• BenchLM's ranking depends on its weights, cohort, and BenchAlign version; category rankings and overall score are not natural constants.

• The Evidence for most Qwen rows points to official Qwen release notes, and BenchLM does not claim on this page to have independently rerun them; the document should therefore be called a verifiable aggregated ledger, not an independent re-run.

• The best verified comparisons shown on the page come from other models currently in the directory; the gaps will change as new models or sources are added.

• When first-party API pricing is unavailable, BenchLM displays N/A; this page's overall score cannot replace testing cost, speed, or context stability.

Reproduction steps

1. Fix the 2026-08-17 data refresh point and save the Qwen3.8 Max model page and the BenchLM methodology page.

2. Export the 52 source-displayable rows to a table, retaining five columns: benchmark, score, weight, cohort, evidence, and original URL.

3. Return to the official Qwen pages for key rows and check the model version, harness, temperature, repetition count, max tokens, and official footnotes.

4. Rerun Coding-, Agentic-, Multimodal-, and HLE-style tasks on your own real task set, and list the new results separately as an independent reproduction rather than replacing BenchLM's aggregate values.

Source excerpt or observation (short quote for compliance only)

BenchLM's evidence constraint for this page is “exact-source rows only”; this makes it suitable as a source ledger and also means it must not be described as a new independent score run.