Configuration: HyInv exact FP32 path (timing configuration) (HyInv report) Protocol: Latency of a design request on one NVIDIA H20 (HyInv report) Dataset: Timing backbones for the HyInv inference-time measurements 130 Runtime (Median wall-clock time of one design request on one NVIDIA H20, 128 candidates per request with a maximum candidate batch of 32 unless the request is the 10,000-design library; the timed region excludes model loading, CUDA graph capture and kernel compilation; the backbone length is named per result; the request is 128 designs for a 148-residue backbone)millisecond · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source Latency of a design request for hyinv-exact-fp32 on one NVIDIA H20 (HyInv report)
hyinv-20261106-protocol-design-latency-h20
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1, row 'HyInv, exact FP32 path', column '128 designs (ms), 148 res.' (page 5 of 21) Configuration: HyInv exact FP32 path (timing configuration) (HyInv report) Protocol: Latency of a design request on one NVIDIA H20 (HyInv report) Dataset: Timing backbones for the HyInv inference-time measurements 303 Runtime (Median wall-clock time of one design request on one NVIDIA H20, 128 candidates per request with a maximum candidate batch of 32 unless the request is the 10,000-design library; the timed region excludes model loading, CUDA graph capture and kernel compilation; the backbone length is named per result; the request is 128 designs for a 402-residue backbone)millisecond · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source Latency of a design request for hyinv-exact-fp32 on one NVIDIA H20 (HyInv report)
hyinv-20261106-protocol-design-latency-h20
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1, row 'HyInv, exact FP32 path', column '128 designs (ms), 402 res.' (page 5 of 21) Configuration: HyInv exact FP32 path (timing configuration) (HyInv report) Protocol: Latency of a design request on one NVIDIA H20 (HyInv report) Dataset: Timing backbones for the HyInv inference-time measurements 14.5 Runtime (Median wall-clock time of one design request on one NVIDIA H20, 128 candidates per request with a maximum candidate batch of 32 unless the request is the 10,000-design library; the timed region excludes model loading, CUDA graph capture and kernel compilation; the backbone length is named per result; the request is a library of 10,000 designs for a 250-residue backbone)second · lower
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source Latency of a design request for hyinv-exact-fp32 on one NVIDIA H20 (HyInv report)
hyinv-20261106-protocol-design-latency-h20
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1, row 'HyInv, exact FP32 path', column '10,000 designs (s)' (page 5 of 21) Configuration: HyInv exact FP32 path (timing configuration) (HyInv report) Protocol: Speed-up of a design request over ProteinMPNN on one NVIDIA H20 (HyInv report) Dataset: Timing backbones for the HyInv inference-time measurements 9.6× Speedup (Printed speed-up against the released ProteinMPNN implementation under the report's timing protocol on one NVIDIA H20; ProteinMPNN's own cell is printed as a dash, so the baseline has no value)unitless · higher
Uncertainty: Not reported by the source
Coverage: Not reported scored / Not reported eligible
Author-reported evaluation · Source checked Evidence concern: excluded from comparisons Appendix C states 'We do not claim zero sequence overlap between the training data and the evaluation panels', and Appendix G's homology screen is a coverage rule under which partial hits remain (Table 10: 28 of 74 de novo backbones and 7 of 12 pocket backbones have partial hits, and all 4 named enzymes are homologous to training data by design). Results on these panels are therefore not clean held-out measurements and must not be compared with results from sources that do establish no overlap.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix C 'Data and pool', final bullet; Appendix G 'Homology of the panels' and Table 10
The ten baselines are single published settings at one temperature each, so their design-set diversity differs from HyInv's: on CATH4.3 test HyInv samples at diversity 0.331 against 0.093 for PiFold, 0.104 for CMLM, 0.144 for LM-Design 650M and 0.000 for DPLM-2 3B, which decodes by argmax. Appendix D.2 states that these four differences are not at matched diversity, and Appendix A states that even for the closest baselines the comparisons are 'close to, but not exactly at, equal diversity', the largest offset of 0.016 being worth at most 0.0011 scTM at the development slope. The scTM and scTM difference results of the protocols named here must not be read as a matched-diversity ranking. The results read at a stated diversity level (Tables 8 and 9 and Appendix D.3) are not covered by this entry, because they are the ones the report reads on a fitted curve at a fixed diversity; nor are the recovery, diversity, pLDDT, scRMSD and lDDT results, whose protocols carry the same single-setting caveat in their own limitations and whose metric qualifiers say the value is at each method's own setting.HyInv Technical Report (Hyra Team, 2026-10-14) · Appendix D.1 and D.2; Table 7; Figure 14 diversity block; Section 5.1 'Baselines' and 'HyInv'
The inference-time results were measured with the supervised-stage weights and a rewrite temperature of 0.23, not with the released checkpoint and its default temperatures, and the quality read-outs of the two numerical paths used the RL checkpoint at update 288, not update 640. Speed and quality in this report are therefore measured on different weights, and the 57x figure compares HyInv's optimised BF16 path with ProteinMPNN's released implementation rather than with an equally optimised baseline.HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1 footnote; Appendix F 'Timing protocol' and 'Baseline timings'
Methods, coverage and source Speed-up of hyinv-exact-fp32 over ProteinMPNN on one NVIDIA H20 (HyInv report)
hyinv-20261106-protocol-design-speedup-vs-proteinmpnn-h20
Aggregation: Not reported
HyInv Technical Report (Hyra Team, 2026-10-14) · Table 1, row 'HyInv, exact FP32 path', column 'Speed-up vs ProteinMPNN' (page 5 of 21)