Atlas

Compare Models

DeepSeek V4.1 Flash 1017 vs Hy4

Hy4 leads by 1.1 capability points (≈95% interval −1.6 to +3.9): statistically tied · P(higher) ≈ 79%

Wins 13 of 19 shared benchmarks (6 lost)

Capability vs. relative cost
DeepSeek V4.1 Flash 1017Hy4Closed weightsOpen weights

Hy4 has no setting with a relative cost yet.

Benchmark wins6 – 13 · 19 shared

Comparing matching setups only (159 benchmarks excluded) · suites counted once
← DeepSeek V4.1 Flash 1017Hy4 →
Data: Surge AI, Mercor, Neel Nanda and 7 othersatlas.village.dev

Margins

Each benchmark scaled 0–100% (its declared range, else the field’s range); Δ is Hy4 minus DeepSeek V4.1 Flash 1017.

RiemannBench (November 2026 rescoring)Math38.451.2+12.8
EnterpriseBench: CoreCraftTool & computer use › Customer service & operations58.568.2+9.7
LLM Debate BenchmarkLanguage & writing › Language & wordplay15791647+9.7
DAXBench v2Software › Data & ML77.787.2+9.6
NoCoT-Bench NCRI 15.2Knowledge & reasoning › Abstract reasoning81.694.0+6.9
NoCoT-Bench NCKI (kspine v3)Knowledge & reasoning › General knowledge102111+6.8
τ³-Banking (Mercor run, bm25_grep retrieval)Tool & computer use › Customer service & operations41.648.1+6.5
CWE-bench v2 Pass@1Software › Cybersecurity57.664.1+6.5
APEX-Agents 1.1Professional › General knowledge work79.485.1+5.7
Harvey LAB (Mercor run, undisclosed ~213-task cohort)Professional › Legal18.813.6−5.2
Agents' Last Exam v1.1 (ALE v1.1)Professional › General knowledge work36.141.0+4.9
Design Arena Overall Code CategoriesSoftware › Frontend & design12891260−4.8
AutomationBench 1.0.6 (provider-reported)Tool & computer use › Customer service & operations60.556.2−4.3
DAYJOB: Finance (October 2026 rescoring)Professional › Finance, accounting & tax12.59.2−3.3
Terminal-Bench 4.0Software › Software engineering42.344.4+2.1
DAYJOB: Healthcare (October 2026 rescoring)Health & medicine6.07.2+1.2
ProgramBench - almost solved rollout percentageSoftware › Software engineering29.130.0+0.9
GDP.xlsxProfessional › General knowledge work18.417.7−0.7
HANDBOOK.md Agents v1.1Tool & computer use › Customer service & operations24.023.7−0.3
Data: Surge AI, Mercor, Neel Nanda and 7 othersatlas.village.dev
Win counts use each benchmark’s declared score direction. Represented suite slices and the derived Capability score are excluded from headline totals. Missing or incompatible setups are excluded, never counted as losses.
Loading Atlas data…