launch80:~$ ls directory/
The run directory.
# benchmark runs and weight scans their publishers chose to list. self-reported unless marked verified. pick up to 4 to compare side by side. how verification works
Add your runs
Benchmark your rig with L80 and publish the result: it shows up here for the community to search and compare. Add --private to keep a page to its link. No install? Publish with curl.
$ curl -fsSL https://github.com/launch80/L80-CLI/releases/download/v1.4.1/install.sh | sh64 of 64 runs
- /QualityBench/full run/Self-reported/by trailfeatures
DeepSeek-V4.1-Flash IQ2_XXS experts · 8x V100-SXM2 PP2×TP4, all experts resident · PLI Volta-sPLIt on 1Cat-vLLM · 150 W · full tier, thinking off vs on
PLI Labs Volta-sPLIt on 1Cat 899e9ed33 (59 patches, MRV2), DSpark 5 · 8x V100-SXM2-32GB PP2xTP4Mean task accuracy 79.9% - /Benchmark report/Self-reported/by jynxter
Qwen3.8-Flash-Next-AA Radiance 2.0.0 TP4 4 × R9700
radiance-2.0.0 · 4x Radeon AI PRO R9700Weighted decode 340.1 t/sTTFT p50 43 ms - /QualityBench/full run/Self-reported/by trailfeatures
DeepSeek-V4.1-Flash IQ2_XXS experts · 4x V100-SXM2 TP4, keep-host expert offload · 1Cat-vLLM v1.5.2-arm1-ds41keep-spc · 150 W · full tier
1Cat-vLLM v1.5.2-arm1-ds41keep-spc (lever1110), DSpark 5 · 4x V100-SXM2-32GB TP4Mean task accuracy 79.9% - /QualityBench/full run/Self-reported/by trailfeatures
Qwen3.8-27B Unsloth UD-Q4_K_M GGUF · 4x V100-SXM2 TP4 · 1Cat-vLLM v1.5.2-arm1 · full tier
1Cat-vLLM v1.5.2-arm1, DFlash2 (7 draft tokens) for tasks; spec off for logprobs · 4x V100-SXM2-32GB TP4Mean task accuracy 75.8%Perplexity 6.50 - /QualityBench/full run/Self-reported/by trailfeatures
Swift-1.5 27B INT4 (W4A16 GPTQ g128) · 4x V100-SXM2 TP4 · 1Cat-vLLM v1.5.2-arm1 · full tier
1Cat-vLLM v1.5.2-arm1, DFlash2 (7 draft tokens) for tasks; spec off for logprobs · 4x V100-SXM2-32GB TP4Mean task accuracy 75.4%Perplexity 6.76 - /BetterBench/quick run/Self-reported/by trailfeatures
DeepSeek-V4.1-Flash IQ2_XXS experts · 4x V100-SXM2 TP4, expert offload · 1Cat-vLLM v1.5.2-arm1 · 150 W
1Cat-vLLM v1.5.2-arm1-ds41keep-spc · 4x Tesla V100-SXM2-32GBWeighted decode 91.1 t/sTTFT p50 651 msPrefill (deepest) 2,136 t/s - /BetterBench/quick run/Self-reported/by trailfeatures
DeepSeek-V4.1-Flash IQ2_XXS experts · 8x V100-SXM2 PP2×TP4, all experts resident · 1Cat-vLLM pli-sp0059 · 150 W
1Cat-vLLM 899e9ed33 + PLI patch series (pli-sp0059) · 8x Tesla V100-SXM2-32GBWeighted decode 131.3 t/sTTFT p50 473 msPrefill (deepest) 2,242 t/s - /BetterBench/quick run/Self-reported/by trailfeatures
Swift-1.5 Flash-Next NVFP4 (FP8 PLE) · 4x V100-SXM2 TP4 · 1Cat-vLLM c59b3c4-arm1b · MTP4 · 4K prefill batch · 300 W
1Cat-vLLM c59b3c4-arm1b, MTP k=4, max-num-batched-tokens 4096 · 4x Tesla V100-SXM2-32GBWeighted decode 140.0 t/sTTFT p50 195 msPrefill (deepest) 4,534 t/s - /Benchmark report/Self-reported/by jynxter
glm-5.3-flash-local Radiance 1.17.0 TP4 fp8 KV 4 × R9700
radiance-pr266-671acfaWeighted decode 67.8 t/sTTFT p50 531 ms - /Benchmark report/Self-reported/by jynxter
deepseek-v4.1-flash-unc-local Radiance 1.17.0 TP4 bf16 KV 4 × R9700
radiance-pr266-671acfaWeighted decode 70.1 t/sTTFT p50 131 ms - /BetterBench/quick run/Self-reported/by jynxter
Qwen3.8-Flash-Next on 4x R9700: one TP4 engine (repeat run)
radiance-1.9.0 · Radeon AI PRO R9700Weighted decode 293.2 t/sTTFT p50 53 msPrefill (deepest) 8,429 t/sAggregate (max conc.) 977.2 t/s - /BetterBench/quick run/Self-reported/by jynxter
Qwen3.8-Flash-Next on 4x R9700: two TP2 copies, random routing
radiance-1.9.0 · Radeon AI PRO R9700Weighted decode 207.9 t/sTTFT p50 65 msPrefill (deepest) 6,645 t/sAggregate (max conc.) 856.0 t/s - /BetterBench/quick run/Self-reported/by jynxter
Qwen3.8-Flash-Next on 4x R9700: two TP2 copies, least-busy routing
radiance-1.9.0 · Radeon AI PRO R9700Weighted decode 212.0 t/sTTFT p50 58 msPrefill (deepest) 6,710 t/sAggregate (max conc.) 955.1 t/s - /BetterBench/quick run/Self-reported/by jynxter
Qwen3.8-Flash-Next on 4x R9700: one TP4 engine
radiance-1.9.0 · Radeon AI PRO R9700Weighted decode 289.2 t/sTTFT p50 52 msPrefill (deepest) 8,495 t/sAggregate (max conc.) 964.6 t/s - /BetterBench/smoke run/Self-reported/by jynxter
Q38NF Radiance 1.9.0 TP4 fp8 KV 4 × R9700
Weighted decode 295.5 t/sTTFT p50 49 msPrefill (deepest) 8,504 t/sAggregate (max conc.) 1,100.7 t/s - /QualityBench/Self-reported/by michaellaunch80
Qwen/Qwen3-1.7B: QualityBench quick run
vllm · RTX 3090Mean task accuracy 54.2%Perplexity 14.21 - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash INT4 baked v2.0 on Arc Pro B65 (quick)
vllm-0.30.0-xpu · Arc Pro B65 32GBWeighted decode 72.5 t/sTTFT p50 232 msPrefill (deepest) 850 t/sAggregate (max conc.) 214.1 t/s - /QualityBench/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash: BF16 vs INT4 v1.0 and v2.0 quality study
vLLM 0.30.0 · RTX PRO 6000Mean task accuracy 76.6%Perplexity 5.98Mean KLD 0.0327 natsTop-1 agreement 93.7% - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash INT4 + DFlash2 k=7 + GDN no-sync patch on Arc Pro B65 (quick)
vllm-0.30.0-xpu · Arc Pro B65 32GBWeighted decode 96.0 t/sTTFT p50 211 msPrefill (deepest) 879 t/sAggregate (max conc.) 203.9 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash INT4 v1.0 + stock DFlash2 k=7 on Arc Pro B65 (quick)
vllm-0.30.0-xpu · Arc Pro B65 32GBWeighted decode 89.3 t/sTTFT p50 213 msPrefill (deepest) 879 t/sAggregate (max conc.) 183.4 t/s - /QualityBench/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash: BF16 vs INT4 v1.0 quality study
vLLM 0.30.0 · RTX PRO 6000Mean task accuracy 76.6%Perplexity 5.98Mean KLD 0.0412 natsTop-1 agreement 92.8% - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash INT4 baked (tool-call calibrated) on Arc Pro B65 (quick)
vllm-0.30.0-xpu · Arc Pro B65 32GBWeighted decode 75.8 t/sTTFT p50 216 msPrefill (deepest) 856 t/sAggregate (max conc.) 210.9 t/s - /QualityBench/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash: BF16 vs INT4 quality study
vLLM 0.30.0 · RTX PRO 6000Mean task accuracy 76.6%Perplexity 5.98Mean KLD 0.0409 natsTop-1 agreement 92.7% - /QualityBench/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash: BF16 vs INT4 quality study
vLLM 0.30.0 · RTX PRO 6000Mean task accuracy 76.6%Perplexity 5.98Mean KLD 0.0409 natsTop-1 agreement 92.7% - /QualityBench/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash: BF16 vs INT4 quality study
vLLM 0.30.0 · RTX PRO 6000Mean task accuracy 76.6%Perplexity 5.98Mean KLD 0.0409 natsTop-1 agreement 92.7% - /QualityBench/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash: BF16 vs INT4 quality study
vLLM 0.30.0 · RTX PRO 6000Mean task accuracy 76.6%Perplexity 5.98Mean KLD 0.0409 natsTop-1 agreement 92.7% - /QualityBench/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash: BF16 quality reference
vLLM 0.30.0 · RTX PRO 6000Mean task accuracy 76.6%Perplexity 5.98 - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwopus3.8-27B-Coder-Flash INT4 baked on Arc Pro B65 (quick)
vllm-0.30.0-xpu · Arc Pro B65 32GBWeighted decode 78.9 t/sTTFT p50 216 msPrefill (deepest) 857 t/sAggregate (max conc.) 219.0 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwen3.5-9B UD-Q4_K_XL on Arc B580 (llama.cpp SYCL) — BetterBench quick
llama.cpp-b11723-sycl · arc-b580-12gbWeighted decode 55.4 t/sTTFT p50 224 msPrefill (deepest) 1,778 t/sAggregate (max conc.) 117.5 t/s - /pwdet scan/Verified/by michaellaunch80
pwdet: unsloth/Qwen3-1.7B — clean · deep
CLEAN · deepChecks passed 70Warnings 6Failures 0 - /pwdet scan/Verified/by michaellaunch80
pwdet: unsloth/Qwen3-0.6B — clean · deep
CLEAN · deepChecks passed 70Warnings 5Failures 0 - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwen/Qwen3-0.6B
Weighted decode 448.7 t/sTTFT p50 393 msPrefill (deepest) 3,720 t/sAggregate (max conc.) 1,936.5 t/s - /BetterBench/quick run/Verified/by michaellaunch80
Launch80/Swift-1.5-Qwen3.8-27B-Uncensored-Int4-baked on vLLM 0.27.2-xpu
vLLM 0.27.2-xpu · Arc Pro B65Weighted decode 79.8 t/sTTFT p50 220 msPrefill (deepest) 857 t/sAggregate (max conc.) 223.4 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Swift-1.5 Qwen3.8-27B Uncensored Int4 baked on 2x H200 NVL (Vast.ai, vLLM TP=2, MTP k=6): quick BetterBench
vllm · 2x H200 NVLWeighted decode 292.4 t/sTTFT p50 449 msPrefill (deepest) 6,903 t/sAggregate (max conc.) 1,208.4 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Swift-1.5 Uncensored Int4 baked + DFlash2 INT4 drafter on Arc Pro B65: quick BetterBench
vllm-0.30.0-xpu · Arc Pro B65 32GBWeighted decode 96.7 t/sTTFT p50 215 msPrefill (deepest) 881 t/sAggregate (max conc.) 196.8 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Swift-1.5 Uncensored Int4 baked on Arc Pro B65: quick BetterBench, MTP k=6
vllm-0.30.0-xpu · Arc Pro B65 32GBWeighted decode 80.5 t/sTTFT p50 214 msPrefill (deepest) 858 t/sAggregate (max conc.) 231.9 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwen3.8-27B TriplePoint Uncensored Int4 baked · Arc Pro B65 · MTP k=6 · INT8 embedding loader on · quick decode-only
Weighted decode 76.7 t/sTTFT p50 216 ms - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwen3.8-27B TriplePoint Uncensored Int4 baked · Arc Pro B65 · MTP k=3 vs k=6 · quick decode-only
Weighted decode 68.7 t/sTTFT p50 208 ms - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwen3.8-27B TriplePoint Uncensored Int4 baked · Arc Pro B65 · MTP k=6 · quick run
vllm-0.30.0-xpu · Arc Pro B65 32GBWeighted decode 77.1 t/sTTFT p50 210 msPrefill (deepest) 856 t/sAggregate (max conc.) 232.4 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
davetha Huihui-Qwen3.8-27B-abliterated AutoRound-Int4 baked · Arc Pro B65 · MTP k=6 · quick run
vllm-0.30.0-xpu · Arc Pro B65 32GBWeighted decode 76.5 t/sTTFT p50 209 msPrefill (deepest) 856 t/sAggregate (max conc.) 220.7 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Swift-1.5 Qwen3.8-27B baked v1 · Arc Pro B65 · vLLM 0.30.0 · XPU graph capture OFF (A/B arm) · MTP k=6 · quick
vLLM 0.30.0 XPU + parity patches (MTP nightly/boundary, GDN v5), R6 draft vocab 32768 · Arc Pro B65 32GBWeighted decode 89.9 t/sTTFT p50 219 msPrefill (deepest) 857 t/sAggregate (max conc.) 247.4 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Swift-1.5 Qwen3.8-27B baked v1 · Arc Pro B65 · vLLM 0.30.0 · XPU graph capture ON (A/B arm) · MTP k=6 · quick
vLLM 0.30.0 XPU + parity patches (MTP nightly/boundary, GDN v5), R6 draft vocab 32768 · Arc Pro B65 32GBWeighted decode 89.0 t/sTTFT p50 211 msPrefill (deepest) 858 t/sAggregate (max conc.) 238.2 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Swift-1.5 Qwen3.8-27B GPTQ-Int4 baked v1 (Wrapzii) · Arc Pro B65 · vLLM 0.30.0 XPU · MTP k=6 · quick run
vLLM 0.30.0 XPU + parity patches (MTP nightly/boundary, GDN v5), R6 draft vocab 32768 · Arc Pro B65 32GBWeighted decode 88.6 t/sTTFT p50 220 msPrefill (deepest) 857 t/sAggregate (max conc.) 247.5 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwen3.8-27B-NVFP4 · 2× Radeon AI PRO R9700 · MXFP4 at load · DFlash2
v0.9.3 · 2x Radeon AI PRO R9700Weighted decode 214.9 t/sTTFT p50 51 msPrefill (deepest) 5,184 t/sAggregate (max conc.) 612.8 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
unsloth/Qwen3.8-27B-NVFP4 · 2× Radeon AI PRO R9700 · MXFP4 at load · DFlash2
v0.9.3 · 2x Radeon AI PRO R9700Weighted decode 214.9 t/sTTFT p50 51 msPrefill (deepest) 5,184 t/sAggregate (max conc.) 612.8 t/s - /BetterBench/full run/Self-reported/by michaellaunch80
Qwen3.8-27B baked v2 · Arc Pro B65 · XPU graph capture OFF (A/B arm) · MTP k=6
Weighted decode 69.0 t/sTTFT p50 215 msPrefill (deepest) 1,043 t/sAggregate (max conc.) 191.9 t/s - /BetterBench/full run/Self-reported/by michaellaunch80
baked v2 · XPU graph capture ON (A/B arm) · Qwen3.8-27B · Arc Pro B65 · MTP k=6
Weighted decode 69.1 t/sTTFT p50 230 msPrefill (deepest) 1,040 t/sAggregate (max conc.) 193.7 t/s - /BetterBench/full run/Self-reported/by michaellaunch80
Project 608 · baked checkpoint v2 (MikeCaldera body, INT4 lm_head + MTP draft on disk) · Qwen3.8-27B · Arc Pro B65
Weighted decode 71.3 t/sTTFT p50 214 msPrefill (deepest) 1,042 t/sAggregate (max conc.) 189.5 t/s - /BetterBench/full run/Self-reported/by michaellaunch80
Project 608 · baked checkpoint (GPTQ lm_head + MTP draft on disk) · Qwen3.8-27B · Arc Pro B65 · MTP k=6
Weighted decode 69.9 t/sTTFT p50 214 msPrefill (deepest) 1,042 t/sAggregate (max conc.) 191.6 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Intel Arc B65 Performance - After Initial R&D
Weighted decode 69.5 t/sTTFT p50 237 msPrefill (deepest) 1,041 t/sAggregate (max conc.) 61.7 t/s - /BetterBench/quick run/Self-reported/by michaellaunch80
Intel Arc B65 Performance - After Initial R&D
- /BetterBench/quick run/Self-reported/by michaellaunch80
B65 R3-03 · Bandwidth cross-check · 434 GB/s sustained
Weighted decode 31.7 t/sTTFT p50 188 ms - /BetterBench/quick run/Self-reported/by michaellaunch80
B65 R3-01 · Target lm_head INT4 · 66.2 t/s decode
Weighted decode 66.2 t/sTTFT p50 241 ms - /BetterBench/quick run/Self-reported/by michaellaunch80
B65 Parity 02 · MTP4 · 48.4 t/s decode
Weighted decode 48.4 t/sTTFT p50 277 ms - /BetterBench/quick run/Self-reported/by michaellaunch80
B65 Parity 01 · No-Spec · 27.9 t/s decode
Weighted decode 27.9 t/sTTFT p50 177 ms - /BetterBench/quick run/Self-reported/by michaellaunch80
Intel Arc B65 Performance - 24.3 t/s decode (OutOfTheBox Config)
Weighted decode 24.3 t/sTTFT p50 194 msPrefill (deepest) 1,077 t/sAggregate (max conc.) 203.2 t/s - /BetterBench/full run/Self-reported/by michaellaunch80
Intel Arc B65 Performance - 24.3 t/s decode (OOTB [BEFORE] Config)
- /BetterBench/full run/Self-reported/by michaellaunch80
SergiioB/Qwen3.8-27B-GPTQ-Int4-sym-G128-MTP-BF16
- /BetterBench/quick run/Self-reported/by michaellaunch80
mlx-community/Qwen3.8-27B-oQ6
Weighted decode 7.3 t/sTTFT p50 1,641 ms - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwen3.8-MXFP4
Weighted decode 218.5 t/sTTFT p50 54 ms - /BetterBench/quick run/Self-reported/by michaellaunch80
Qwen3.8-MXFP4
Weighted decode 218.5 t/sTTFT p50 54 ms - /Benchmark report/Self-reported/by michaellaunch80
mlx-community/Qwen3.8-27B-oQ6
Weighted decode 7.3 t/sTTFT p50 1,641 ms