localairecipes launch configs for local inference

Leaderboard

Sort by Then by

NVIDIA DGX Spark 128GB ×4 (Blackwell)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
GLM-5.3-Flash NVFP4 4x DGX Spark switchless ring TP4 DFlash2 #1	42	45.6 tok/s	110.8 tok/s	—	36074.9 tok/s	1817 ms	reported	0
DeepSeek-V4.1-Flash on 4x DGX Spark — tuned TP4 SGLang/DSpark	39	89.7 tok/s	165.9 tok/s	244.6 tok/s	5664 tok/s	—	reported	0
GLM-5.3-Flash exlr8 EXL3 experts on DGX Spark TP1-4	n/a	128 tok/s	—	—	4393 tok/s	—	reported	0

NVIDIA DGX Spark 128GB ×3 (Blackwell)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
GLM-5.3-Flash exlr8 EXL3 experts on DGX Spark TP1-4 #1	n/a	128 tok/s	—	—	2687 tok/s	—	reported	0

NVIDIA DGX Spark 128GB ×2 (Blackwell)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
Qwen3.8-Flash-Next Uncensored NVFP4 2x DGX Spark SGLang TP2 ROCE #1	40	40.3 tok/s	117.3 tok/s	120.5 tok/s	2976.7 tok/s	—	reported	0
GLM-5.3-Flash exlr8 EXL3 experts on DGX Spark TP1-4	n/a	128 tok/s	—	—	4393 tok/s	—	reported	0

NVIDIA DGX Spark 128GB ×1 (Blackwell)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
GLM-5.3-Flash VQ Q4_K 1x DGX Spark 1M ctx (JLFN) #1	42	24.3 tok/s	12.1 tok/s	—	—	—	reported	0
Qwen3.8-Flash-Next NVFP4 DGX Spark GB10 SGLang NEXTN MTP	40	110.4 tok/s	—	—	—	465 ms	reported	0
Qwen3.8-Flash-Next W4A16-AutoRound DGX Spark vLLM MTP4 Iter7g	40	76.8 tok/s	179.9 tok/s	261.2 tok/s	—	—	reported	0
Qwen3.8-Flash-Next NVFP4 1x DGX Spark vLLM 71.3 c1 (salient-data)	40	69.3 tok/s	—	91.9 tok/s	—	—	reported	0
Qwen3.8-Flash-Next FP4-Local 1x DGX Spark vLLM b12x	40	66.2 tok/s	166.4 tok/s	—	—	42.2 ms	reported	0
Qwen3.8-Flash-Next Uncensored NVFP4 BF16 PLE DGX Spark vLLM MTP2	40	44.2 tok/s	—	—	2055.2 tok/s	116800 ms	reported	0
Qwen3.8-Flash-Next NVFP4 DGX Spark GB10 SGLang MTP-213 262K	40	32.7 tok/s	93.2 tok/s	—	—	286 ms	reported	0
Qwen3.8-27B SGLang NVFP4 DFlash2 on 1x DGX Spark (deterministic)	34	68.3 tok/s	81.7 tok/s	141.5 tok/s	—	—	reported	0
Nemotron 3.5 Bolt 74B-A9B 1x DGX Spark vLLM b12x LiLiCorr 262K	n/a	140 tok/s	113.4 tok/s	163.6 tok/s	—	—	reported	0
GLM-5.3-Flash exlr8 EXL3 experts on DGX Spark TP1-4	n/a	128 tok/s	—	—	4393 tok/s	—	reported	0

NVIDIA dgx-station-gb300 288GB ×2 (Blackwell)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
DeepSeek V4.1 Flash 1017 FP8 TP1xPP2 DSpark 2x DGX Station GB300 RoCE vLLM 0.31 overlay #1	45	391 tok/s	595 tok/s	974 tok/s	—	—	reported	0

NVIDIA RTX 5090 32GB ×2 (Blackwell)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
GLM-5.3-Flash EXL3 2.05bpw on 2x RTX 5090 - TabbyAPI hot-expert swaps, 262K ctx, vision #1	42	62.9 tok/s	98 tok/s	—	2023 tok/s	480 ms	reported	0

NVIDIA RTX PRO 6000 Blackwell 96GB ×4

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
GLM-5.3-Flash on 4x RTX PRO 6000 Blackwell \x{2014} TP4 MTP3 #1	42	248.9 tok/s	—	873.8 tok/s	13514 tok/s	—	reported	0
GLM-5.3-Flash on 4x RTX PRO 6000 Blackwell \x{2014} TP4 DFlash2 K7	42	219.7 tok/s	—	677.7 tok/s	13840 tok/s	—	reported	0
DeepSeek V4.1 Flash on 4x RTX PRO 6000 Blackwell \x{2014} TP4 DSpark K7	39	230.6 tok/s	—	748.3 tok/s	17503 tok/s	—	reported	0

NVIDIA RTX PRO 6000 Blackwell 96GB ×3

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
DeepSeek-V4.1-Flash-1017 Kraken TP3 3x RTX PRO 6000 Blackwell #1	45	271.8 tok/s	474.3 tok/s	—	13486 tok/s	—	reported	0

NVIDIA RTX PRO 6000 Blackwell 96GB ×2

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
GLM-5.3-Flash QAD (CSF) on 2x RTX PRO 6000 Blackwell — TP2/DCP2 MTP3 #1	42	207 tok/s	488 tok/s	672 tok/s	—	—	reported	0
DeepSeek V4 Flash Vision on 2x RTX PRO 6000 Blackwell \x{2014} TP2 DSpark K3	35	169.2 tok/s	394.7 tok/s	—	8993 tok/s	—	reported	0
DeepSeek V4 Flash-0731 on 2x RTX PRO 6000 Blackwell \x{2014} TP2 DSpark K5	34	214.9 tok/s	—	611.9 tok/s	11167 tok/s	—	reported	0

NVIDIA RTX PRO 6000 Blackwell 96GB ×1

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
GLM-5.3-Flash EXL3 3.05bpw on 1x RTX PRO 6000 - GPU expert cache, host-RAM streaming #1	42	68.3 tok/s	—	—	2851.4 tok/s	—	reported	0
Qwen3.8-Flash-Next NVFP4 RTX PRO 6000 Blackwell SM120 SGLang	40	227.8 tok/s	—	—	14235.1 tok/s	10101 ms	reported	0
Qwen3.8-Flash-Next NVFP4 RTX PRO 6000 Blackwell SGLang NVMe PLE	40	180.9 tok/s	408 tok/s	—	9990 tok/s	92 ms	reported	0
Qwen3.8-Flash-Next NVFP4 on RTX PRO 6000 Blackwell — TP1 MTP3	40	157.6 tok/s	—	674.2 tok/s	12104 tok/s	—	reported	1
Qwen3.8-27B NVFP4A4-AWQ 1x RTX PRO 6000 SGLang Fast Decode	34	149 tok/s	117 tok/s	40 tok/s	—	—	reported	0
Qwen3.8-27B Bit-Exact Spec-Decode 1x RTX PRO 6000 SGLang	34	106.6 tok/s	—	—	—	68500 ms	reported	0

NVIDIA CMP 170HX 64GB ×6 (Ampere)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
DeepSeek-V4.1-Flash SM80 DeepInfer kernel port 6x CMP 170HX DSpark K5 #1	39	160.8 tok/s	170.5 tok/s	—	4506.8 tok/s	10700 ms	reported	0

NVIDIA CMP 170HX 64GB ×4 (Ampere)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
DeepSeek-V4.1-Flash-1017 Engram-NVMe TP4/PP4 4x CMP 170HX #1	45	355.7 tok/s	—	867.6 tok/s	2416 tok/s	—	reported	0
DeepSeek-V4.1-Flash-1017 Hybrid EXL3 TP4 4x CMP 170HX Morrowmake	45	319 tok/s	—	765 tok/s	3818 tok/s	—	reported	0
DeepSeek-V4.1-Flash-1017 EXL3 2.9bpw TC-MoE on 4x CMP 170HX TP4	45	195.4 tok/s	—	276.6 tok/s	1391.1 tok/s	—	reported	0
DeepSeek-V4.1-Flash-1017 Fidelity-First 4x CMP 170HX alanoo81	45	105.4 tok/s	325.2 tok/s	505 tok/s	2514 tok/s	—	reported	0
GLM-5.3-Flash on 4x CMP 170HX — Morrowmake recipe 1.8.0, TP4 DFlash2	42	264.4 tok/s	—	—	6264 tok/s	63 ms	reported	1
Qwen3.8-Flash-Next-FP8 4x CMP 170HX — PP4 MTP-4 patched vLLM (g4bwy)	40	184.4 tok/s	548 tok/s	—	10138 tok/s	91900 ms	reported	0
DeepSeek-V4.1-Flash EXL3 3bpw TC-MoE on 4x CMP 170HX TP4 (original checkpoint)	39	178.8 tok/s	343.9 tok/s	586.6 tok/s	1570 tok/s	—	reported	0
DeepSeek-V4.1-Flash (764B) on 4x CMP 170HX — PP4, shadow KV, EXL3 2bpw, DSpark K5, 512k ctx, native vision (kaka86mm)	39	60.7 tok/s	—	200.6 tok/s	—	—	reported	0
MiMo-V2.6-Flash-RL 309B MoE NVFP4 4x CMP 170HX TP4 DFlash 1M ctx (ElectroHobo)	38	322 tok/s	—	753.8 tok/s	2722 tok/s	—	reported	0
MiMo-V2.6-Flash-RL (omni) on 4x CMP 170HX — PP4, official fp8, three-tier KV, v3.3 (kaka86mm)	38	64.5 tok/s	—	—	6050 tok/s	3300 ms	reported	0

NVIDIA CMP 170HX 64GB ×3 (Ampere)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
Qwen3.8-Flash-Next W4A16 3x CMP 170HX PP3 NVMe PLE fp8 KV 1M (clee) #1	40	74.9 tok/s	—	—	4690 tok/s	223000 ms	reported	1
DeepSeek V4 Flash-0731 abliterated on 3x CMP 170HX — vLLM patched PP3 DSpark K5	34	181 tok/s	—	—	2350 tok/s	—	reported	0

NVIDIA CMP 170HX 64GB ×2 (Ampere)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
Qwen3.8-Flash-Next EXL3 5.05bpw 2x CMP 170HX TP2 MTP3 (Morrowmake 1.0.0) #1	42	152.8 tok/s	458.6 tok/s	674.7 tok/s	2651 tok/s	—	reported	0
GLM-5.3-Flash EXL3 3.05bpw on 2x CMP 170HX TP2 DFlash2 K7 hybrid drafting 393K (alexvnesta)	42	123.6 tok/s	115 tok/s	—	1316 tok/s	—	reported	0
GLM-5.3-Flash low effort on 2x CMP170HX EXL3 TP2 four slots	42	122.8 tok/s	155.5 tok/s	—	—	325.7 ms	reported	0
GLM-5.3-Flash high effort on 2x CMP170HX EXL3 TP2 four slots	42	113 tok/s	143.7 tok/s	—	—	314.4 ms	reported	0
GLM-5.3-Flash EXL3 3.05bpw on 2x CMP 170HX — DFlash2 K7 Q8 KV 384K (Flun)	42	90 tok/s	—	—	—	—	reported	1
GLM-5.3-Flash max effort on 2x CMP170HX EXL3 TP2 four slots	42	49.8 tok/s	101.2 tok/s	—	—	331.8 ms	reported	0
Qwen3.8-Flash-Next W4A16 2x CMP 170HX TP2 DFlash2 host-RAM PLE	40	348.9 tok/s	—	992.1 tok/s	4729 tok/s	—	reported	0
Qwen3.8-Flash-Next-Uncensored AWQ W4A16 on 2x CMP 170HX — 8 slots, PP2, MTP5 + int4 draft head, PLE-from-SSD (mannydelrio)	40	240.4 tok/s	608.1 tok/s	1044 tok/s	6964 tok/s	—	reported	0
Qwen3.8-Flash-Next W4A16 2x CMP 170HX PP2 MTP5 PLE (AlexBrandes)	40	226.2 tok/s	560.2 tok/s	840.4 tok/s	7096 tok/s	80 ms	reported	0
Qwen3.8-Flash-Next INT4-Mixed 170HX x2, PP2 MTP3 BF16-KV PLE-NVMe (bozocentral)	40	199 tok/s	183 tok/s	—	—	—	reported	0
Qwen3.8-Flash-Next NVFP4 2x CMP 170HX TEP2 GDS SSD PLE MTP6	40	170.8 tok/s	—	—	2083.1 tok/s	—	reported	0
Qwen3.8-Flash-Next W4A16 INT4PLE on 2x modded CMP 170HX — PP2 MTP3 (DUFRelic)	40	133.4 tok/s	245.2 tok/s	—	6213.7 tok/s	—	reported	0
Qwen3.8-Flash-Next W4A16-FP8PLE 2x CMP 170HX PP2 MTP4 1M	40	112 tok/s	—	—	6080 tok/s	47400 ms	reported	0
Qwen3.8-Flash-Next-Uncensored AWQ W4A16 on 2x CMP 170HX — PP2, MTP3, PLE-from-SSD (mannydelrio)	40	109 tok/s	283 tok/s	—	7700 tok/s	13000 ms	reported	0
Qwen3.8-Flash-Next W4A16 2x CMP 170HX TP2+EP MTP4 FP8 PLE 92GB Host	40	104.5 tok/s	392 tok/s	—	2839 tok/s	70500 ms	reported	0

NVIDIA CMP 170HX 64GB ×1 (Ampere)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
Qwen3.8-Flash-Next 3bpw FP8-PLE 1x CMP 170HX DFlash2 RecoverSSM #1	40	296.4 tok/s	—	799.5 tok/s	4243 tok/s	—	reported	0
Qwen3.8-Flash-Next 3bpw MTP on 1x CMP 170HX SSD PLE AIO (iIIusi0n)	40	111 tok/s	78 tok/s	61 tok/s	2636 tok/s	105 ms	reported	0
Qwen3.8-27B INT8 MTP3 prefix-cache on 1x CMP 170HX Hynix	34	79 tok/s	—	—	—	—	reported	0

NVIDIA RTX 3090 24GB ×4 (Ampere)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
Qwen3.8-Flash-Next W4A16-Merlin on 4x RTX 3090 — TP2xPP2+EP, MTP3, FP8 KV, 925K pool (halt95) #1	40	174.3 tok/s	—	533.7 tok/s	5484 tok/s	960 ms	reported	0
Qwen3.8-27B AWQ-INT4 + DSpark on RTX 3090s — SGLang TP1/TP2/TP4	34	136.8 tok/s	103.8 tok/s	60.5 tok/s	1615.8 tok/s	98880 ms	reported	0

NVIDIA RTX 3090 24GB ×2 (Ampere)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
DeepSeek-V4.1-Flash-1017 dsx 2.6-bit GPTQ 2x RTX 3090 (Engram) #1	45	26 tok/s	—	—	170 tok/s	—	reported	0
GLM-5.3-Flash EXL3 3.05bpw on 2x RTX 3090 — vLLM + ExLlamaV3 CPU expert offload, 256K (DominikBucko)	42	50.4 tok/s	—	—	1155 tok/s	235260 ms	reported	0
Qwen3.8-Flash-Next W4A16-FP8PLE on 2x RTX 3090 — TP2, expert-offload, MTP3 (DominikBucko)	40	109.8 tok/s	—	—	3029 tok/s	90800 ms	reported	0
Qwen3.8-27B AWQ-INT4 + DSpark on RTX 3090s — SGLang TP1/TP2/TP4	34	136.8 tok/s	103.8 tok/s	60.5 tok/s	1615.8 tok/s	98880 ms	reported	0

NVIDIA RTX 3090 24GB ×1 (Ampere)

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
GLM-5.3-Flash EXL3 3.05bpw on 1x RTX 3090 — RAM/NVMe expert tiering (near-exact) #1	42	28.2 tok/s	—	—	1015 tok/s	—	reported	0
Qwen3.8-Flash-Next W4A16-FP8PLE on 1x RTX 3090 — CPU cold experts, MTP3, 128K (DominikBucko)	40	50.3 tok/s	—	—	2141 tok/s	62200 ms	reported	0
Qwen3.8-27B AWQ-INT4 + DSpark on RTX 3090s — SGLang TP1/TP2/TP4	34	136.8 tok/s	103.8 tok/s	60.5 tok/s	1615.8 tok/s	98880 ms	reported	0

AMD Radeon AI Pro R9700 32GB ×2

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
Qwen3.8-27B PARO-MXFP6 2x Radeon AI PRO R9700 — vLLM radiance TP2 #1	34	138 tok/s	321.6 tok/s	394 tok/s	3570 tok/s	—	reported	0

AMD Radeon AI Pro R9700 32GB ×1

Recipe	AA intel	C1 decode	C4 decode	C8 decode	Prefill	TTFT	Evidence	Votes
Qwen3.8-27B MXFP4 DFlash2 1x Radeon AI PRO R9700 WSL2 radiance #1	34	145 tok/s	—	—	2762 tok/s	12580 ms	reported	0

Machine-readable: /api/v1/leaderboard.json (these boards) · /api/v1/scores.json (per-recipe medians).