Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++
27.8 KB
Raw Blame

Community bench results, index

One row per community folder in this directory (ported for 0.1.40.2 and 0.1.41 from the community bench PRs, authors credited in the commits). Raw dumps, logs and binaries were trimmed from the folders; each trimmed folder has a TRIMMED.md listing what was left out. The README in each folder has the full method.

How to read the table:

  • Prompt is prompt (prefill) tok/s; Output is decode tok/s. Cells with / give the 4K / 32K / 128K runs unless the cell says otherwise.
  • Numbers are the submitter's own medians, taken from each folder's README. Different machines, prompts and engine versions: compare rows with care. Our own numbers for the same engine are in the dated reports above (2026-09-29-speed-0126 and others).
  • Strata is the engine version the submitter ran.
  • PR is the original pull request (closed by the 2026-10-06 history rewrite; the work is in 0.1.40.2).
Folder Hardware Model Prompt tok/s Output tok/s Strata PR
2026-10-01-0.1.31-release RX 7900 XTX 24 GB, Ryzen 7 7700X Coder IQ1_M 545 to 900 median (0.1.30 to 0.1.31) 16-34 to 17-65 0.1.30 / 0.1.31 #404
2026-10-01-community-2x-rtx-pro-4500 2x RTX PRO 4500 32 GB, Threadripper 7960X Swift IQ3_XXS 1,428 / 2,454 / 2,785 115 / 120 / 93 0.1.30 #418
2026-10-02-community-2x-rtx-pro-4500-engine-0.1.36 2x RTX PRO 4500 32 GB, Threadripper 7960X Swift IQ3_XXS 2,960 / 5,065 / 5,762 127 / 130 / 105 0.1.36 #418
2026-10-01-community-2x-titan-rtx 2x TITAN RTX 24 GB (NVLink), Xeon E5-2696 v4 IQ3_S, 262K 865 / 1,612 / 1,785 68 / 60 / 62 0.1.31 #389
2026-10-01-community-rtx-pro-4500 RTX PRO 4500 32 GB, Threadripper PRO 3975WX UD-Q4_K_XL / UD-IQ4_XS / IQ3_S see README UD-Q4_K_XL 61.5 / 72.1 / 70.2; UD-IQ4_XS 82.0 / 95.8 / 91.4; IQ3_S 106 / 120 / 117 (2.7K / 64K / 128K) 0.1.31 #417
2026-10-01-community-rtx-5090-5950x RTX 5090, Ryzen 9 5950X (AVX2), 96 GB UD-Q4_K_XL, IQ3_S, Swift IQ3_XXS UD-Q4_K_XL 14.7K: 2,161 (0.1.31) UD-Q4_K_XL 14.7K: 88.3 0.1.31 to 0.1.39 #433
2026-10-01-ctx-ladder RTX 4070 Ti SUPER 16 GB (docker) IQ2_XS 167 / 128 / 133 (SSE prefill, 32K / 64K / 128K) 97 / 84 / 74 (32K / 64K / 128K, KV q4_0 at 128K) see README #995
2026-10-05-community-rtx-4070-ti-super RTX 4070 Ti SUPER 16 GB, Ryzen 7 9800X3D IQ2_XS / IQ3_XXS cold prefill in README (about 2,500 at 32K-256K) IQ2_XS 97.9 / 94.3 / 90.0 / 82.4; IQ3_XXS 76 / 71 / 70 / 56 (32K / 64K / 128K / 256K) 0.1.39 (6f32ec0) #995
2026-10-01-iq4xs-64gb RTX 5060 Ti 16 GB (PCIe 4.0 x8), Ryzen 7 5700X3D, 64 GB UD-IQ4_XS (60.9 GiB of experts) n/a (ms per draft round) 33.9 pinned arena, 16.9-39 mmap variants 0.1.31 / 0.1.32 #416
2026-10-02-community-2x5060ti 2x RTX 5060 Ti 16 GB (PCIe gen3), i9-9900KF IQ3_S 1,108 (20.9K cold) 56.4 (20.9K); 62.9 short 0.1.34 #483
2026-10-02-community-rtx-5060-ti-x2 2x RTX 5060 Ti 16 GB, Xeon E5-2690 v4 IQ3_XXS, 262K 808 / 1,931 / 2,199 77.5 / 72.7 / 72.4 0.1.35 #508
2026-10-05-community-rtx-5060-ti-x2 2x RTX 5060 Ti 16 GB, Xeon E5-2690 v4 IQ3_XXS, 262K (rerun) 911 / 2,087 / 2,440 84.8 / 86.5 / 83.8 0.1.39 #508
2026-10-02-community-rtx-5090-iq3s RTX 5090, Ryzen 9 9950X3D, 96 GB IQ3_S, 262K 2,473 / 4,168 / 4,450 (4K / 32K / 131K) 141.8 / 142.2 / 127.2 0.1.33 #440
2026-10-02-community-rtx2080ti-11gb RTX 2080 Ti 11 GB, Threadripper 3960X, 128 GB IQ3_S, 262K 497 / 637 / 566 39.6 / 38.0 / 34.1 0.1.33 #469
2026-10-02-community-rtx4000ada-layer-split 2x RTX 4000 Ada (+ 1x SFF Ada), EPYC 7313P IQ3_S 2,110 (2 GPU) vs 779 (3 GPU), 16.6K 73 vs 74 median 0.1.35 #521
2026-10-02-community-rtx5090-1m-agent RTX 5090, Ryzen 9 7950X IQ3_S, 1M context (YaRN) about 2,500 / 3,300 / 3,400 (4K / 32K / 128K) 94-105 0.1.31 #466
2026-10-02-community-rx9070xt-windows RX 9070 XT 16 GB, Windows 11 IQ3_S about 600 45-46 (37 first run) 0.1.35 (AMD engine) #499
2026-10-03-ashley-2080ti RTX 2080 Ti, i9-9900KF, Windows 11 IQ3_XXS 525 at 32K 33-38 (1K to 128K) 0.1.38 #698
2026-10-03-dev-5090 RTX 5090, Ryzen 9 9950X3D IQ3_S (0.1.34 / 0.1.38 / STRATA_PF_FUSED=1) 4K: 4,308 / 4,846 / 5,361 4K: 178 / 193 / 189; 128K: 144 / 152 / 165 0.1.34, 0.1.38 #698
2026-10-03-community-r9700-windows Radeon AI PRO R9700 (gfx1201), Ryzen 9 9950X, Windows 11 IQ2_XS 856 (4K) / 1,182 (32K) / 1,232 (131K) / 1,130 (248K) 99.8 (4K) / 97.8 / 89.9 / 87.5 0.1.38 #618
2026-10-03-community-rtx4090-iq3s-140k-code RTX 4090, Ryzen 9 7950X IQ3_S, 140K, code prompts 2,059 / 2,948 / 3,048 77.5 / 97.2 / 84.4 0.1.38 #624
2026-10-03-community-rtx4090-iq3s-140k-ru RTX 4090, Ryzen 9 7950X IQ3_S, 140K, Russian prose 2,084 / 2,991 / 3,078 76.1 / 102.5 / 102.2 0.1.38 #624
2026-10-03-community-rtx4090-iq3xxs-200k-code RTX 4090, Ryzen 9 7950X IQ3_XXS, 200K, code prompts 1,844 / 2,976 / 2,962 118 / 128 / 108 0.1.38 #624
2026-10-03-community-rtx4090-iq3xxs-200k-ru RTX 4090, Ryzen 9 7950X IQ3_XXS, 200K, Russian prose 1,824 / 2,912 / 2,984 94.2 / 85.6 / 90.3 0.1.38 #624
2026-10-04-community-rtx4090-iq3xxs-200k-code RTX 4090, Ryzen 9 7950X IQ3_XXS, 200K, code prompts 1,966 / 2,974 / 3,085 124 / 135 / 120 0.1.39 #624
2026-10-04-community-rtx4090-iq3xxs-200k-ru RTX 4090, Ryzen 9 7950X IQ3_XXS, 200K, Russian prose 1,908 / 3,128 / 3,126 81.1 / 90.5 / 87.8 0.1.39 #624
2026-10-03-rx6900xt-coder RX 6900 XT, Threadripper 3990X (Windows) Coder IQ1_M 313 (4K) / 350 (32K) 45.3 (4K) with 31 workers vs 11.6 with 63 0.1.38 #628
2026-10-03-windows-groups RX 6900 XT, Threadripper 3990X (Windows) Coder IQ1_M see README 12.7 to 46.1 (short), 13.1 to 47.4 (4K) with processor groups kept 0.1.38 + fix #628
2026-10-03-community-rx7900xtx-rocm102-nightly RX 7900 XTX 24 GB, ROCm 7.1.1 vs 10.2 nightly IQ3_S 921 vs 1,535 (without / with matching hipBLASLt) 62.0 vs 61.0 0.1.38 #745
2026-10-04-community-rx7900xtx-hipblaslt-100500 RX 7900 XTX 24 GB, ROCm 10.2 nightly IQ3_S, 128K 926 to 1,687 (gfx1100 table) 62.3 0.1.38 lineage #745
2026-10-03-community-v100 Tesla V100-PCIE 32 GB, 2x Xeon E5-2696 v3, Windows Flash-Next (see README) see README 16.5 to 33.5 after calibration; 22.0 prose 0.1.38, updated for 0.1.39 #707
2026-10-03-community-z840-rtx3090ti RTX 3090 Ti, 2x Xeon E5-2699 v3 (HP Z840) IQ3_S, 262K 1,239-1,304 (about 5K) 90-99 (NUMA variants) 0.1.38 #674
2026-10-04-community-2x-rtx3060 2x RTX 3060 12 GB, Xeon E5-2678 v3 IQ3_S 603 / 1,156 / 1,323 (3.2K / 24.8K / 98.6K) 43.6 / 42.5 / 40.0 0.1.37 (source) #721
2026-10-04-community-rx-7900-gre RX 7900 GRE 16 GB, Ryzen 7 5700X3D IQ2_XS / IQ3_XXS / Coder IQ1_M 1,140-1,216 (8.8K) 65.3 / 59.1 / 54.7 0.1.38 #740
2026-10-04-community-mi50 MI50 (gfx906), i5-12400F IQ2_XS about 330 / 330 / 305 36.0 / 34.0 / 33.5 0.1.38 #758
2026-10-04-community-rtx-4090-laptop RTX 4090 Laptop, i9-13900HX IQ3_S 1,345 / 2,506 / 2,386 51.2 / 49.9 / 48.1 0.1.38 #757
2026-10-04-community-rtx-5090-9700x RTX 5090, Ryzen 7 9700X, Windows 11 IQ3_XXS 3,775 (median, fresh) 171.7 0.1.39 #777
2026-10-05-community-rtx-5090-9700x-iq3s RTX 5090, Ryzen 7 9700X, Windows 11 IQ3_S 3,072 (median, fresh) 164.3 0.1.39 #811
2026-10-04-community-rtx-4080s-iq3s RTX 4080 SUPER, i7-13790F, Windows 11 IQ3_S cold 2,535 (pool 4) vs 2,486 (pool 15) 95 (4 workers) vs 69-78 (15 workers) mean 0.1.39 #780
2026-10-04-community-rx-6800-windows RX 6800, Windows Flash-Next (see README) 326 (7.9K) / 311 (30.3K) 42.6 to 56.0 English (0.1.39 vs STRATA_SH_STREAM=0) 0.1.38 / 0.1.39 #815
2026-10-04-community-v100-16gb-ram Tesla V100-PCIE 32 GB (x4 link), i7-13700KF, 16 GB RAM Coder IQ1_M 1,233 / 1,580 / 1,394 69.3 / 68.6 / 67.1 0.1.39 #823
2026-10-04-community-rtx-5070ti-windows RTX 5070 Ti, Ryzen 7 9800X3D, Windows 11 IQ3_S 3,012 (5.2K) / 3,487 (41K) / 3,523 (131K) / 3,442 (164K) 75.8 / 74.7 / 82.8 / 79.6 0.1.39 #832
2026-10-04-community-4090-32gb-ram-512k RTX 4090 24 GB, 32 GB RAM IQ2_XS, 512K 3,411 (477K prompt) 122.0 at depth 0.1.39 #834
2026-10-04-community-v100-rtx4070-split Tesla V100 32 GB + RTX 4070 12 GB, 16 GB RAM IQ2_XS 985 / 1,535 / 1,533 79.1 / 83.8 / 77.1 0.1.39 #850
2026-10-04-community-4090-agents-vs-dense-27b RTX 4090 24 GB, 32 GB RAM IQ2_XS vs a dense 27B (coding-agent task) about 3,800 (70-110K re-reads) 173-185 0.1.39 + 4 PRs #882
2026-10-05-community-v100-sxm2 Tesla V100-SXM2 32 GB, Windows (CUDA 12.6) Q2_0 / UD-IQ4_XS / third quant (see README) Q2_0: 1,268 (8K) Q2_0 76.4; UD-IQ4_XS 26.3 0.1.39 (6f32ec0) #902
2026-10-05-community-e5-2673v3-rtx-4060ti RTX 4060 Ti 16 GB, Xeon E5-2673 v3 (DDR3, PCIe 3.0 x8) Q2_0 / IQ3_S (two arms each) 715 (Q2_0) / 439 (IQ3_S), 3.5K Q2_0 48.1 to 55.7; IQ3_S 30.5 to 35.0 0.1.39 #913
2026-10-06-community-rtx5070ti-5900x RTX 5070 Ti, Ryzen 9 5900X, Windows 10 IQ2_XS / IQ3_XXS / IQ3_S IQ2_XS 1,992 / 2,817; IQ3_XXS 1,653 / 2,482; IQ3_S 1,439 / 2,261 IQ2_XS 102 / 101; IQ3_XXS 72.5 / 73.7; IQ3_S 62.9 / 66.1 0.1.39 #993
2026-10-06-community-rtx-5060ti RTX 5060 Ti 16 GB, Ryzen 9 7945HX IQ3_S 1,158 / 1,225 / 1,176 56.6 / 60.2 / 57.2 0.1.35 (d9ab843) #1016
2026-10-05-community-rx-6700-xt RX 6700 XT 12 GB (gfx1031), Ryzen 7 5700G Flash-Next (see README) 302 / 309 / 298 32.1 / 30.7 / 30.2 0.1.39 #1027
2026-10-05-community-2x-p40 2x Tesla P40 22 GB, 2x Xeon E5-2698 v4 IQ2_XS 353 (4K) / 501 (16K), both cards; 348 / 388 one card 34.8 both cards, 19.6 one card 0.1.39 #1028
2026-10-05-community-2x-rtx-4060ti 2x RTX 4060 Ti 16 GB, Threadripper PRO 3975WX UD-IQ4_XS 880 / 1,917 / 2,221 54.4 / 56.6 / 54.7 0.1.39 #1062
2026-10-06-community-rx-6800m RX 6800M (gfx1031), Windows, self-built HIP Flash-Next quants (see README) 50-114 9-27 0.1.40 #1078
2026-10-05-community-2x-rx-6900xt 2x RX 6900 XT 16 GB (gfx1030), Ryzen 5 5600X, 128 GB IQ3_S, 131K one card stock 457 / 474 / 459, with #835 + #849 + #854 868 / 1,100 / 1,023; layer split with them 853 / 1,616 / 1,823 40-43 one card, 57-61 layer split 0.1.39 stock and with #835, #849, #854 #927
2026-10-05-community-gfx1151 Radeon 8060S (gfx1151), Ryzen AI Max+ 395, 128 GB unified memory Q2_0, UD-IQ4_XS 250.9 to 528.9 with the hipBLASLt table (matched synthetic prompts) 45.6 to 46.8 (shared-expert stream on) 0.1.39 + #895, #820 #917
2026-10-05-community-arc-b65 Intel Arc Pro B65 32 GB (SYCL), Core i5-12600K, 128 GB IQ2_XS, 8K 307-359 (512 and 7,000-token tasks) 36-49 by task; medians of the five tasks 40.4-41.3 / 39.8-40.9 0.1.40 source + 6 local patches #955

Notes:

  • #927 ran seven configurations; the table shows the one-card stock and PR sets and the layer split with the PRs (#835 is opt-in in 0.1.40, so those are not the release default), and its README has the helper-mode rows. #917 is 0.1.39 plus then-unmerged #895 and #820, and 0.1.40 turned the shared-expert stream off on HIP (#826). #955's patches 3, 5 and 6 are source changes in 0.1.40.2.
  • #1016 ran its engine at the 0.1.35 commit and the table shows the numbers as submitted.
  • #995 also changed the canonical 2026-09-29-speed-0126 report (matrix.json rewritten, README extended with another host's rows); that part was not ported, only the two new folders.
  • #777 and #811 are the same machine on two quants, kept as two folders.
  • The earlier community folders (2026-09-30-community-rtx-5090, 2026-10-03-community-2x-mi50, 2026-10-04-community-2x-arc-pro-b60) are listed in docs/COMMUNITY_BENCHMARKS.md. | 2026-10-06-community-rtxpro-v0140 | RTX PRO 6000 | Q4 / Q8 (MTP and ngram runs) | see the README | see the README | 0.1.40 | #1137 | | 2026-10-06-community-2x-rx-7900-gre | 2x RX 7900 GRE (gfx1100) | layer split, #848 / #854 | see the README | see the README | 0.1.40 | #1140 | | 2026-10-06-community-2x-p100 | 2x Tesla P100 16 GB | Flash-Next IQ3_S, 128K | see the README | see the README | 0.1.40 | #1157 | | 2026-10-06-community-rtx4090-opt-ins | RTX 4090 | IQ3_S 140K / IQ3_XXS 200K opt-ins and tool-call hotfix | see the README | see the README | 0.1.38 to 0.1.40 | #1158 | | 2026-10-06-community-rtx4090-toolcall-hotfix | RTX 4090 | IQ3_S 140K / IQ3_XXS 200K opt-ins and tool-call hotfix | see the README | see the README | 0.1.38 to 0.1.40 | #1158 | | 2026-10-06-q8-prefill8192-rtxpro | RTX PRO 6000 | Q8, --prefill 8192 | see the README | see the README | 0.1.40 | #1171 | | 2026-10-06-community-rx-7900-xtx | RX 7900 XTX, Windows 11 | IQ3_S | see the README | see the README | 0.1.40 | #1173 | | 2026-10-06-community-4x-tesla-p100 | 4x Tesla P100 16 GB (CUDA 12 engine) | IQ3_XXS | see the README | see the README | 0.1.40 | #1192 | | 2026-10-06-community-rtx3090-egpu-64gb | RTX 3090 eGPU, 64 GB RAM | IQ2_XS vs IQ3_XXS | see the README | see the README | 0.1.40 | #1199 | | 2026-10-06-community-r9-2x-rtx-2080-ti | 2x RTX 2080 Ti (R9) | see README | see the README | see the README | 0.1.40 | #1225 | | 2026-10-06-community-2x-rtx-pro-4500-engine-0.1.40.1 | 2x RTX PRO 4500 Blackwell | Swift IQ3_XXS | see the README | see the README | 0.1.40.1 | #1226 | | 2026-10-05-community-rtx5080-9900x3d | RTX 5080, Ryzen 9 9900X3D | Coder IQ1_M and others | see the README | see the README | 0.1.40 | #1243 | | 2026-10-07-community-rtx-5090-laptop-ud-iq4xs | RTX 5090 Laptop 24 GB | UD-IQ4_XS (first NVIDIA measurement) | see the README | see the README | 0.1.40.1 | #1245 | | 2026-10-06-community-rtx-5090-laptop | RTX 5090 Laptop 24 GB, Windows 11 | see README | see the README | see the README | 0.1.40.1 | #1263 | | 2026-10-06-community-2x-rx-6900xt-0.1.40.1 | 2x RX 6900 XT 16 GB (gfx1030), Ryzen 5 5600X, 128 GB | IQ3_S, 131K | one card stock 458 / 473 / 459, with the two gfx103x switches 811 / 973 / 913, with #1150 + #1151 + #1167 969 / 1,152 / 1,126; layer split with them 968 / 1,685 / 2,003 | one card 47.5-51.8; layer split stock 62.1 / 67.0 / 59.0, with the PRs 72.2 / 74.7 / 70.3; helper 64-75 | 0.1.40.1 stock, the two switches, and with #1150, #1151, #1167 | #1270 | | 2026-10-07-community-arc-b65-v01402 | Arc Pro B65 32 GB (Gen4 x16), i5-12600K | Flash-Next IQ2_XS, 8K / 262K profiles | 7K input: 315 (8K, prefill 512), 830 (prefill 4096), 288 (262K) | 43.5 (8K), 42.9 (prefill 4096), 42.4 (262K) at 7K | 0.1.40.2 (SYCL) | #1432 | | 2026-10-07-community-rtx5090-pro4000-helper | RTX 5090 32 GB + RTX PRO 4000 Blackwell 24 GB (helper cache), Ryzen 7 9800X3D, Windows 11 | UD-Q4_K_XL, 262K | 1,514 / 2,048 / 2,201 (5K / 33K / 115K prompt) | 103 / 92 / 94 | 0.1.40.3 | #1433 | | 2026-10-07-community-2x-rtx-4000 | 2x Quadro RTX 4000 8 GB, 2x Xeon E5-2620 v3, Proxmox LXC | Swift IQ2_XS, 131K | 99.9 to 249.2 on 73K-token prompts (fresh and cached tail) | 18.1 to 22.8 | 0.1.38 | #1443 | | 2026-10-07-community-v100-p100-helper | Tesla V100 32 GB + Tesla P100 16 GB (P100 as expert helper, mixed sm_70 + sm_60), Xeon E5-2699 v3 | Flash-Next IQ3_XXS, 262K | 1,437 (18K) / 1,446 (44K) / 1,411 (120K, one run) | about 73 (256-token cap) | 0.1.40.2 | #1452 | | 2026-10-07-community-xtx-6800xt-1m-helper | RX 7900 XTX 24 GB + RX 6800 XT 16 GB (helper), Xeon E5-2696 v4, 1M context, source HIP build with local patches | UD-Q4_K_XL | 1,350 (50K) / 1,294 (130K) / 592 (1,048K cold) | 97 long coding (final arm); 39 at the full 1M depth | 0.1.40.2 + local patches | #1459 | | 2026-10-07-community-rtx-5080-docker-wsl2 | RTX 5080 16 GB in Docker Desktop on Windows (WSL2), Ryzen 9 9950X | Coder IQ1_M, 131K and 65K | 2,253 / 3,304 (4K / 32K, 131K limit) | 75.7 / 80.8 / 72.2 (4K / 32K / 130K) | 0.1.40.3 | #1460 | | 2026-10-08-community-rtx-5090-laptop-engine-0.1.40.3 | RTX 5090 Laptop 24 GB (175 W), Ryzen 9 9955HX3D, Windows 11 | Flash-Next Q2_0, 131K | 1,950 / 2,893 / 2,780 (4K / 32K / 128K) | 138 / 144 / 131 | 0.1.40.3 (0.1.40.2 in a short section) | #1462 | | 2026-10-08-community-rtx4090-iq3xxs-200k-code | RTX 4090, Ryzen 9 7950X, 46 GB RAM, Linux source build | Flash-Next IQ3_XXS, 204,800, code prompts, experimental speed projection on | 2,146 / 3,342 / 3,292 (4K / 32K / 125K) | 128 / 138 / 122 | 0.1.40.3 | #1467 | | 2026-10-08-community-rtx4090-iq3xxs-200k-ru | RTX 4090, Ryzen 9 7950X (same PC) | Flash-Next IQ3_XXS, 204,800, Russian prompts | see the README | see the README | 0.1.40.3 | #1467 | | 2026-10-08-community-2x-titan-rtx-0.1.40.3 | 2x TITAN RTX 24 GB (NVLink, unused), Xeon E5-2696 v4, source build | Flash-Next IQ3_S, 262K | 869 / 1,549 / 1,627 (4K / 32K / 128K) | 70.9 / 77.3 / 71.1 | 0.1.40.3 | #1429 | | 2026-10-07-community-gfx1150 | Radeon 890M (gfx1150), Ryzen AI 9 HX PRO 370, 96 GB unified memory | IQ3_XXS | 153 / 216 / 226 (1K / 3.6K / 7K) with the gfx1150 hipBLASLt table, 99 / 124 / 130 without | 15.2-17.6 | 0.1.40.2 | - | | 2026-10-12-community-gfx1103 | Radeon 780M (gfx1103), Ryzen 7 8745HS, 60 GB, GTT 50 GiB, hand build on ROCm 10.1 | Flash-Next Q2_0 + MTP, 32K | 90 plain / 186 with the gfx1103 hipBLASLt table / 211 with the exact switches / 277 with the WMMA opt-ins (4K); 272 at 32K (WMMA arm) | 13.1 plain, 13.9-14.3 tuned | 0.1.43 + the gfx1103 port (branch) | - | | 2026-10-09-community-2x-rx-6900xt-0.1.41 | 2x RX 6900 XT 16 GB (PCIe 4.0 x8 each), Ryzen 5 5600X, 128 GB | IQ3_S, 131K | one card stock 473 / 502 / 486, with the two gfx103x switches 850 / 1,102 / 1,027, with the PRs 976 / 1,174 / 1,147; layer split stock 464 / 767 / 869, switches 832 / 1,619 / 1,823, PRs 968 / 1,714 / 2,029; helper with the PRs 986 / 1,173 / 1,147 | one card 46-52; layer split stock 62.6 / 65.0 / 60.8, PRs 64.7 / 68.8 / 68.3, PRs + --spec 3 --spec-min-p 0.7 --pipeline-windows 2 77.2 / 82.3 / 75.7; helper PRs 72.5 / 75.2 / 75.1 | 0.1.41 (fb58e0db) | this PR; notes: Vop3p/dual-6900xt-flash-next-tuning | | 2026-10-09-community-gfx1151-ud-iq4xs-0.1.41 | EVO-X2, Radeon 8060S (gfx1151), Ryzen AI Max+ 395, 128 GB UMA | Unsloth UD-IQ4_XS | 627 / 687 / 617 default; 808 / 1,165 / 1,109 fast | 42.8 / 44.4 / 42.1 default; 40.1 / 49.6 / 47.5 fast | 0.1.41 (fb58e0d) | - | | 2026-10-08-community-2x-rtx3080-20gb-ud-q4-k-xl-512k-pool | 2x RTX 3080 20 GB (220 W cap each, PCIe Gen3 x16), Xeon E5-2696 v4 (AVX2), 90 GiB RAM, Proxmox LXC | Unsloth UD-Q4_K_XL, two 262,144-token lanes in a 524,288-token shared KV pool, layer split, batch MTP | 718 / 2,219 / 2,585 / 2,996 / 3,080 (4K / 25K / 51K / 104K / 127K) | 81.2 / 84.0 / 90.6 / 86.3 / 80.7 (same lengths, one stream); 89.5 for two streams (my script) | fork build of 0.1.40.3, source branch repro/w7 on noon-at-cgn/Strata (not upstream main plus the open PRs: it contains the code of my open PRs #1598, #1599, #1601, #1614 (shared KV pool), #1636 (batch MTP on a split), #1637 (--adapt-async beside slots), and of #1190, an open PR by another contributor that keeps the RAM copy of the split lend regions) | - | | 2026-10-08-community-r9700-linux | 1x Radeon AI PRO R9700 32 GB, 2x EPYC 9334, 503 GiB (Linux, ROCm 7.2.3) | Flash-Next IQ3_S (IQ2_XS contrast), frozen experimental 0.1.40.3 | 633 / 768 / 751 / 718 (1K / 4K / 32K / 131K) | 101.9 / 96.4 / 95.6 / 87.6; short-increment pairs for #1107 | 0.1.40.3 (frozen experimental) | #1638 | | 2026-10-08-community-v100-sm70-table | Tesla V100-SXM2 32 GB, Threadripper 2990WX, 96 GB | UD-IQ4_XS native-experts pack, 262K | unchanged by STRATA_SM70_TABLE=1 | 28.16 / 28.64 to 30.92 / 32.94 (+9.8% / +15.0%, two order-swapped pairs, bit-identical tokens) | 0.1.41 (fb58e0d), source build | #1647 | | 2026-10-09-community-strix-halo-laptop-70w | Radeon 8060S, Ryzen AI Max+ 395 laptop (70 W), 128 GB unified | UD-IQ4_XS, STRIX_HALO.md section 5 fast config | 969 / 1,067 / 1,031 (8K / 32K / 128K) | 59.2 / 57.8 / 42.0 | 0.1.41 | #1716 | | 2026-10-09-community-rtx-5080-4060ti-0.1.41 | RTX 5080 16 GB + RTX 4060 Ti 8 GB (helper cache), Ryzen 9 9950X3D, 96 GB | IQ3_S, 131K (stock v0.1.41; a private patched build and two larger files are listed separately in the README) | 3,451 / 4,277 / 4,166 | 146.0 / 132.2 / 124.9 | 0.1.41 stock | #1681 | | 2026-10-09-community-4x-rx-7900-xt | 4x RX 7900 XT 20 GB (gfx1100), i9-9900K, 60 GB, asymmetric Gen3 PCIe | IQ3_S, 262K, layer split | five opt-ins off 756 / 1,722 / 2,433; on 2,052 / 3,804 / 4,307 | off 59.3 / 52.0 / 49.6; on 60.0 / 52.7 / 50.4 | 0.1.41 | #1655 | | 2026-10-09-community-p4-prefill-roofline | 1x Tesla P4, 2x Xeon E5-2697 v3, 252 GiB | GSQ-RCO IQ3_XXS (original Flash-Next), int8 KV | fixed 512: 78.0 (4K) / 78.8 (8K); existing auto: 107.3 / 114.8 | not measured | 0.1.41 (fb58e0d) | #1695 | | 2026-10-09-rope-yarn-long-context | RTX PRO 6000 Blackwell 96 GB, Ryzen 9 7950X, 124 GiB | ISTA IQ3_XXS, 512K and 1M, ordinary RoPE vs YaRN | 5,479 (512K RoPE) / 4,452 (1M RoPE) / 4,505 (1M YaRN) | 103.6 / 82.4 / 82.7; recall 3/3 (512K RoPE), 0/3 (1M RoPE), 3/3 (1M YaRN) | 0.1.41 (fb58e0d) | #1692 | | 2026-10-09-community-rtx-5080-w7-2475x-ud-iq4xs-786k | RTX 5080 16 GB, Xeon w7-2475X, 128 GB DDR5 (Windows 11) | Unsloth UD-IQ4_XS, 786K context (YaRN x3), needles to 512K | 1,814 / 3,299 / 3,398 (4K / 32K / 128K) | 68.6 / 70.2 / 64.1 | 0.1.41 | #1726 | | 2026-10-09-community-rtx-3060-0141 | RTX 3060 12 GB, Core i7-12700, 64 GB (Windows 11) | Flash-Next Q2_0 / IQ2_XS / IQ3_XXS / IQ3_S, 128 to 260K | Q2_0 882 / 981 / 934 (4K / 32K / 128K); IQ3_XXS 681 / 884 / 867; IQ3_S 569 / 820 / 803 | Q2_0 42.9 / 41.8 / 38.6; IQ3_XXS 26.2 / 27.1 / 27.0; IQ3_S 20.2 / 20.2 / 18.9 | 0.1.41 (official Windows release) | #1725 | | 2026-10-09-community-rtx5090-9800x3d-swift-1m | RTX 5090, Ryzen 7 9800X3D, 91.9 GiB usable RAM, 480 W | Swift IQ3_XXS, 1M YaRN | 6,687 / 6,426 (32K / 131K, greedy, n=2) | 198.9 / 190.6 (same runs); native 1M disk restore 6.107 s | 0.1.41 | #1722 | | 2026-10-10-community-rx-9070 | RX 9070 16 GB (gfx1201), Ryzen 7 5800X3D, 64 GB RAM, Windows 11 | Flash-Next IQ3_XXS, 131K | 586 / 804 / 833 (4.2K / 33.5K / 130K) | 58.7 / 57.9 / 57.9 | 0.1.41 | #1811 | | 2026-10-09-community-rtx-5060-ti-e5-2680-v4 | RTX 5060 Ti 16 GB (PCIe 3.0 x8), Xeon E5-2680 v4, Windows 10 | Coder IQ1_M / abliterated Q2_0, 262K | Coder 1,180 / 1,601 / 1,860; Q2 782 / 1,549 / 1,514 | Coder 47.5 / 51.0 / 49.1; Q2 68.0 / 73.4 / 68.5 | 0.1.41 | #1784 | | 2026-10-09-community-radeon-pro-vii-16gb | Radeon Pro VII 16 GB (gfx906), Xeon E5-2666 v3, host build on the distribution's ROCm 6.2 | Flash-Next IQ2_XS, 131K | 350 / 339 / 284 (4K / 32K / 128K) | 24.2 / 24.0 / 23.7 | 0.1.40.4 (0.1.41 in a second round) | #1793 | | 2026-10-10-community-rx-9060-xt | RX 9060 XT 16 GB (gfx1200), Ryzen 7 9800X3D, 64 GB DDR5, Windows 11 | Flash-Next IQ2_XS / IQ3_XXS, prefill sweep, 32K coding comparison, 113K recall | IQ2_XS 879 / 807 (8K / 28K, prefill 8,192); 790 (113K) | IQ2_XS 56.3 / 57.6 / 46.9 (8K / 28K / 113K); coding weighted decode IQ2_XS 49.6 vs IQ3_XXS 42.5 | 0.1.40.3 (prebuilt Windows HIP) | #1802 | | 2026-10-10-community-2x-p40-iq3_s | 2x Tesla P40 24 GB, Xeon E5-2699 v3, 125 GB RAM, Ubuntu 22.04 | Flash-Next IQ3_S, 131K, layer split | 350 / 558 (3.9K / 21.9K) | 35.0 (short prompt); needles 6/6 at 32K and 128K | 0.1.41 (source build) | #1799 | | 2026-10-09-community-rtx-5060ti-wsl2-iq3_s-0141 | RTX 5060 Ti 16 GB (PCIe 4.0 x8), Ryzen 7 3700X, 128 GB RAM, WSL2 | Flash-Next IQ3_S, 131K, experimental speed projection | 343 / 598 (3.8K / 19.2K), single runs | 48.2 / 47.7 (1,400 / 1,284 tokens), single runs | 0.1.41 | #1751 | | 2026-10-08-community-rtx4070-iq2-xs | RTX 4070 12 GB, Core i5-12600KF, 32 GB DDR4, Windows 11 | Flash-Next IQ2_XS, 100K cells, q4_0 KV, adaptive cache E4/S24 | 28,912-token prompt (not redistributed) | 48.23 (accepted-decode median of 5); a modified, locally built engine, not an official release | modified local build (provenance.json) | #1739 |