OpenRecipes

Tuned and verified LLM training recipes

Throughput-optimized training recipes for frontier open‑weight models, each measured on one fixed SFT workload with its cost per million tokens and a passed convergence check.

8,192-token SFT rows · global batch 128 · BF16

Submit a recipe Protocol GitHub

Recipes

verified

Models

in 10 families

Model–GPU pairs

on H100, H200, B200

Models with recipes

Each GPU row shows its fastest recipe; the bar compares it with the model's fastest.

Muse2 recipes

Muse-Spark-Open-816B-A42B

816B parameters · N 44.4B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H200333.1×128$6.60

Starting point224.4 → 333.1 +48%

B200505.2×128$7.83

Starting point343.3 → 505.2 +47%

Newest 2026-10-17

Qwen4 recipes

Qwen3.8-27B

27B parameters · N 25.6B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H1001,480.4×8$1.29

Starting point861.5 → 1,480.4 +72%

H2001,194.3×8$1.84 B2004,011.4×8$0.99

Starting point1,411.1 → 4,011.4 +184%

Newest 2026-10-18

Qwen2 recipes

Qwen3.6-27B

27B parameters · N 25.6B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H2002,499.5×8$0.88

Starting point948.7 → 2,499.5 +163%

B2004,330.9×8$0.91

Starting point1,434.4 → 4,330.9 +202%

Newest 2026-10-18

Qwen5 recipes

Qwen3.6-35B-A3B

35B parameters · N 2.9B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H2004,663.1×8$0.47 B2006,573.4×8$0.60

Newest 2026-10-14

Qwen2 recipes

Qwen3.5-35B-A3B

35B parameters · N 2.9B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H2004,749.0×8$0.46 B2006,597.0×8$0.60

Newest 2026-10-14

Qwen8 recipes

Qwen3.5-122B-A10B

122B parameters · N 9.0B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100999.7×48$1.91 H2002,451.4×32$0.90 B2002,475.8×16$1.60

Newest 2026-10-18

Qwen9 recipes

Qwen3.5-397B-A17B

397B parameters · N 16.3B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100313.3×128$6.10 H200545.3×128$4.03 B2001,232.5×64$3.21

Newest 2026-10-18

Qwen1 recipe

Qwen3-30B-A3B

30B parameters · N 3.0B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H2006,913.2×8$0.32

Starting point3,129.7 → 6,913.2 +121%

B200No recipe yet

Newest 2026-10-21

Qwen1 recipe

Qwen3-32B

32B parameters · N 32.0B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H2002,200.2×8$1.00

Starting point1,327.8 → 2,200.2 +66%

B200No recipe yet

Newest 2026-10-18

GLM8 recipes

GLM-5.2-744B-A40B

744B parameters · N 40.3B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H10068.3×256$28.00 H200161.8×128$13.58 B200146.8×64$26.95

Newest 2026-10-13

GLM1 recipe

GLM-4.7-Flash

N 3.6B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H2006,178.5×8$0.36

Starting point3,142.9 → 6,178.5 +97%

B200No recipe yet

Newest 2026-10-15

GLM1 recipe

GLM-4.5

N 32.9B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H200565.4×128$3.89

B200No recipe yet

Newest 2026-10-21

Kimi1 recipe

Kimi-K2.6

N 31.7B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H200353.4×256$6.22

B200No recipe yet

Newest 2026-10-22

DeepSeek1 recipe

DeepSeek-V4-Flash

N 13.3B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H200808.2×64$2.72

B200No recipe yet

Newest 2026-10-13

Nemotron1 recipe

Nemotron-3-Nano-30B-A3B

30B parameters · N 3.2B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H2007,814.0×8$0.28

Starting point5,229.2 → 7,814.0 +49%

B200No recipe yet

Newest 2026-10-15

Nemotron1 recipe

Nemotron-3-Super-120B-A12B

120B parameters · N 12.2B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H2001,045.2×16$2.10

Starting point688.3 → 1,045.2 +52%

B200No recipe yet

Newest 2026-10-21

Gemma2 recipes

Gemma-4-26B-A4B

26B parameters · N 3.8B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H2004,023.2×8$0.55

Starting point1,878.2 → 4,023.2 +114%

B2003,763.4×8$1.05

Starting point1,989.9 → 3,763.4 +89%

Newest 2026-10-20

Gemma1 recipe

Gemma-4-31B

31B parameters · N 30.7B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H1001,178.1×8$1.62

Starting point743.4 → 1,178.1 +58%

H200No recipe yet

B200No recipe yet

Newest 2026-10-20

GPT-OSS2 recipes

gpt-oss-120b

120B parameters · N 5.1B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H1002,569.4×64$0.74

H200No recipe yet

B200No recipe yet

Newest 2026-10-14

GPT-OSS6 recipes

gpt-oss-20b

20B parameters · N 3.6B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H1007,496.7×8$0.25 H20010,483.9×8$0.21 B20014,536.3×8$0.27

Newest 2026-10-14

MiniMax1 recipe

MiniMax-M2

N 10.4B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H2002,034.4×64$1.08

Starting point557.8 → 2,034.4 +265%

B200No recipe yet

Newest 2026-10-21

Inkling1 recipe

Inkling-Small

N 11.2B active

GPUvs. fastestTok/s/GPUGPUs$/1M

H100No recipe yet

H200616.7×32$3.56

Starting point396.9 → 616.7 +55%

B200No recipe yet

Newest 2026-10-21

Latest recipes

Newest first; rank is within the model and GPU pair.

Measured	Model	GPU	GPUs	Tok/s/GPU	$ / 1M	Layout	Rank	Author
2026-10-22	Kimi-K2.6	H200	256	353.4	$6.22	TP8 PP16 EP16 DP2PP 16, no recompute, expandable segments off	#1 of 1	ATTUNE
2026-10-21	Qwen3-30B-A3B	H200	8	6,913.2	$0.32	TP1 PP1 EP8 DP8max 16,384 tokens/GPU, no offload, allocator without expandable segments	#1 of 1	ATTUNE
2026-10-21	GLM-4.5	H200	128	565.4	$3.89	TP4 PP16 EP8 DP2PP 16, no recompute	#1 of 1	ATTUNE
2026-10-21	Nemotron-3-Super-120B-A12B	H200	16	1,045.2	$2.10	TP4 PP2 EP8 DP2max 16,384 tokens/GPU, selective MoE recompute, CPU optimizer offload 0.55	#1 of 1	ATTUNE
2026-10-21	MiniMax-M2	H200	64	2,034.4	$1.08	TP1 PP8 VPP4 EP8 DP8VPP 4, no offload	#1 of 1	ATTUNE
2026-10-21	Inkling-Small	H200	32	616.7	$3.56	TP4 PP4 EP8 DP212/10/10/10 layer split, CPU optimizer offload 0.7	#1 of 1	ATTUNE
2026-10-20	Gemma-4-26B-A4B	H200	8	4,023.2	$0.55	TP1 PP1 EP8 DP8selective recompute of attention and MoE, cuDNN fused sliding-window attention, no offload	#1 of 1	ATTUNE
2026-10-20	Gemma-4-26B-A4B	B200	8	3,763.4	$1.05	TP1 PP1 EP8 DP8one row per microbatch, selective recompute, no offload	#1 of 1	ATTUNE
2026-10-20	Gemma-4-31B	H100	8	1,178.1	$1.62	TP2 PP4 DP113/16/16/15 layer split, CPU optimizer offload 0.8, cuDNN fused sliding-window attention	#1 of 1	ATTUNE
2026-10-18	Qwen3.8-27B	H100	8	1,480.4	$1.29	TP1 PP4 DP216/17/17/14 layer split, CPU optimizer offload	#1 of 1	ATTUNE
2026-10-18	Qwen3.8-27B	B200	8	4,011.4	$0.99	TP1 PP4 DP217/17/17/13 layer split, block recompute of 14 layers, no offload	#1 of 2	ATTUNE
2026-10-18	Qwen3.8-27B	B200	8	3,858.5	$1.03	TP1 PP2 DP434/30 layer split, no offload	#2 of 2	ATTUNE
2026-10-18	Qwen3.6-27B	H200	8	2,499.5	$0.88	TP1 PP4 DP216/17/17/14 layer split, block recompute of 12 layers, no offload	#1 of 1	ATTUNE
2026-10-18	Qwen3.6-27B	B200	8	4,330.9	$0.91	TP1 PP87/9/9/9/9/9/9/3 layer split, block recompute of 2 layers, no offload	#1 of 1	ATTUNE
2026-10-18	Qwen3.5-122B-A10B	B200	16	2,475.8	$1.60	TP1 PP4 EP4 DP4max 8,192 tokens/GPU	#1 of 2	ATTUNE
2026-10-18	Qwen3.5-122B-A10B	B200	16	2,447.9	$1.62	TP1 PP8 EP2 DP2max 8,192 tokens/GPU	#2 of 2	ATTUNE
2026-10-18	Qwen3.5-397B-A17B	B200	64	1,232.5	$3.21	TP2 PP8 VPP2 EP8 DP4max 16,384 tokens/GPU, VPP 2, no offload	#1 of 3	ATTUNE
2026-10-18	Qwen3.5-397B-A17B	B200	64	1,130.7	$3.50	TP2 PP8 VPP2 EP8 DP4max 8,192 tokens/GPU, VPP 2, no offload	#2 of 3	ATTUNE
2026-10-18	Qwen3.5-397B-A17B	B200	48	1,075.5	$3.68	TP2 PP6 EP8 DP4max 8,192 tokens/GPU, no offload	#3 of 3	ATTUNE
2026-10-18	Qwen3-32B	H200	8	2,200.2	$1.00	TP4 PP2max 16,384 tokens/GPU, no recompute, no offload	#1 of 1	ATTUNE
2026-10-17	Muse-Spark-Open-816B-A42B	H200	128	333.1	$6.60	TP4 PP16 EP8 DP2max 8,192 tokens/GPU, full recompute, CPU optimizer offload 0.3	#1 of 1	ATTUNE
2026-10-17	Muse-Spark-Open-816B-A42B	B200	128	505.2	$7.83	TP2 PP16 EP8 DP4max 8,192 tokens/GPU, full recompute, Adam on the GPU, cuDNN fused attention	#1 of 1	ATTUNE
2026-10-15	GLM-4.7-Flash	H200	8	6,178.5	$0.36	TP1 PP1 EP8 DP8max 16,384 tokens/GPU, no offload	#1 of 1	ATTUNE
2026-10-15	Nemotron-3-Nano-30B-A3B	H200	8	7,814.0	$0.28	TP1 PP2 EP4 DP4max 8,192 tokens/GPU, no offload	#1 of 1	ATTUNE
2026-10-14	Qwen3.8-27B	H200	8	1,194.3	$1.84	TP4 PP2max 8,192 tokens/GPU, no recompute, CPU optimizer offload	#1 of 1	ATTUNE
2026-10-14	Qwen3.6-35B-A3B	B200	8	6,573.4	$0.60	TP1 PP1 EP8 DP8max 16,384 tokens/GPU, optimizer on GPU	#1 of 2	ATTUNE
2026-10-14	Qwen3.5-35B-A3B	B200	8	6,597.0	$0.60	TP1 PP1 EP8 DP8max 16,384 tokens/GPU, optimizer on GPU	#1 of 1	ATTUNE
2026-10-14	gpt-oss-120b	H100	64	2,569.4	$0.74	TP1 PP8 EP8 DP8PP 8 with uneven split 3/5/5/5/5/5/5/3, full recompute	#1 of 2	ATTUNE
2026-10-14	gpt-oss-120b	H100	64	2,326.7	$0.82	TP1 PP4 EP8 DP16uneven pipeline split 9/10/10/7, full recompute	#2 of 2	ATTUNE
2026-10-14	gpt-oss-20b	H100	8	7,496.7	$0.25	TP1 PP2 EP4 DP413/11 layer split, block recompute of 8 layers	#1 of 2	ATTUNE
2026-10-14	gpt-oss-20b	H200	8	10,483.9	$0.21	TP1 PP1 EP4 DP8no recompute	#1 of 2	ATTUNE
2026-10-14	gpt-oss-20b	B200	8	14,536.3	$0.27	TP1 PP1 EP8 DP8max 16,384 tokens/GPU, selective moe_act recompute	#1 of 2	ATTUNE
2026-10-13	Qwen3.6-35B-A3B	H200	8	4,663.1	$0.47	TP1 PP1 EP8 DP8max 8,192 tokens/GPU, no offload	#1 of 3	ATTUNE
2026-10-13	Qwen3.6-35B-A3B	H200	8	3,466.8	$0.63	TP1 PP1 EP8 DP8max 16,384 tokens/GPU	#2 of 3	ATTUNE
2026-10-13	Qwen3.6-35B-A3B	H200	8	3,076.8	$0.71	TP1 PP2 EP4 DP4max 16,384 tokens/GPU	#3 of 3	ATTUNE
2026-10-13	Qwen3.6-35B-A3B	B200	8	4,334.1	$0.91	TP1 PP1 EP8 DP8max 8,192 tokens/GPU, no offload	#2 of 2	ATTUNE
2026-10-13	Qwen3.5-35B-A3B	H200	8	4,749.0	$0.46	TP1 PP1 EP8 DP8max 8,192 tokens/GPU, no offload	#1 of 1	ATTUNE
2026-10-13	Qwen3.5-122B-A10B	H100	48	842.8	$2.27	TP4 PP6 EP8 DP2max 8,192 tokens/GPU, flex dispatcher	#2 of 3	ATTUNE
2026-10-13	Qwen3.5-122B-A10B	H100	48	822.1	$2.32	TP4 PP3 EP16 DP4max 8,192 tokens/GPU	#3 of 3	ATTUNE
2026-10-13	Qwen3.5-122B-A10B	H200	32	2,451.4	$0.90	TP1 PP4 EP8 DP8max 8,192 tokens/GPU, recompute 2 layers, load-balancing loss off	#1 of 3	ATTUNE
2026-10-13	Qwen3.5-122B-A10B	H200	32	2,289.7	$0.96	TP1 PP4 EP8 DP8max 8,192 tokens/GPU, flex dispatcher, load-balancing loss off	#3 of 3	ATTUNE
2026-10-13	Qwen3.5-397B-A17B	H100	128	313.3	$6.10	TP2 PP16 VPP2 EP8 DP4max 8,192 tokens/GPU, VPP 2, CPU optimizer offload	#1 of 3	ATTUNE
2026-10-13	Qwen3.5-397B-A17B	H100	128	288.6	$6.62	TP2 PP8 VPP2 EP16 DP8max 8,192 tokens/GPU, VPP 2, CPU optimizer offload	#3 of 3	ATTUNE
2026-10-13	Qwen3.5-397B-A17B	H200	128	506.7	$4.34	TP1 PP4 EP32 DP32max 8,192 tokens/GPU, no offload	#2 of 3	ATTUNE
2026-10-13	Qwen3.5-397B-A17B	H200	128	469.2	$4.68	TP2 PP2 EP64 DP32max 8,192 tokens/GPU, no offload	#3 of 3	ATTUNE
2026-10-13	GLM-5.2-744B-A40B	H100	256	68.2	$28.03	TP4 PP8 EP32 DP8max 8,192 tokens/GPU, recompute 2 layers, CPU optimizer offload	#2 of 3	ATTUNE
2026-10-13	GLM-5.2-744B-A40B	H100	256	66.2	$28.85	TP4 PP8 CP2 EP32 DP4max 8,192 tokens/GPU, CPU optimizer offload	#3 of 3	ATTUNE
2026-10-13	GLM-5.2-744B-A40B	H200	128	131.1	$16.76	TP8 PP4 EP32 DP4max 8,192 tokens/GPU, selective recompute, CPU optimizer offload	#2 of 2	ATTUNE
2026-10-13	GLM-5.2-744B-A40B	B200	64	146.8	$26.95	TP4 PP2 CP4 EP32 DP2max 16,384 tokens/GPU, CPU optimizer offload (fraction 0.9)	#1 of 3	ATTUNE
2026-10-13	GLM-5.2-744B-A40B	B200	64	136.5	$28.99	TP4 PP2 CP8 EP32 DP1max 8,192 tokens/GPU, CPU optimizer offload (fraction 0.9)	#3 of 3	ATTUNE
2026-10-13	DeepSeek-V4-Flash	H200	64	808.2	$2.72	TP2 PP4 EP8 DP8one row per microbatch, no offload	#1 of 1	ATTUNE
2026-10-13	gpt-oss-20b	H100	8	6,736.9	$0.28	TP1 PP2 EP4 DP4one row per microbatch, full recompute	#2 of 2	ATTUNE
2026-10-13	gpt-oss-20b	H200	8	9,448.2	$0.23	TP1 PP2 EP4 DP4one row per microbatch, no recompute	#2 of 2	ATTUNE
2026-10-13	gpt-oss-20b	B200	8	12,566.1	$0.31	TP1 PP2 EP4 DP4one row per microbatch, no recompute	#2 of 2	ATTUNE
2026-10-11	Qwen3.5-122B-A10B	H100	48	999.7	$1.91	TP4 PP6 EP8 DP2max 16,384 tokens/GPU, flex dispatcher	#1 of 3	ATTUNE
2026-10-11	Qwen3.5-122B-A10B	H200	32	2,366.7	$0.93	TP1 PP4 EP8 DP8max 8,192 tokens/GPU, load-balancing loss off	#2 of 3	ATTUNE
2026-10-11	Qwen3.5-397B-A17B	H100	128	301.9	$6.33	TP2 PP16 EP8 DP4max 8,192 tokens/GPU, CPU optimizer offload	#2 of 3	ATTUNE
2026-10-11	Qwen3.5-397B-A17B	H200	128	545.3	$4.03	TP2 PP4 EP32 DP16max 8,192 tokens/GPU, no offload	#1 of 3	ATTUNE
2026-10-11	GLM-5.2-744B-A40B	H100	256	68.3	$28.00	TP4 PP8 EP32 DP8max 8,192 tokens/GPU, CPU optimizer offload	#1 of 3	ATTUNE
2026-10-11	GLM-5.2-744B-A40B	H200	128	161.8	$13.58	TP4 PP4 EP32 DP8max 8,192 tokens/GPU, selective recompute, CPU optimizer offload	#1 of 2	ATTUNE
2026-10-11	GLM-5.2-744B-A40B	B200	64	146.7	$26.97	TP4 PP2 CP4 EP32 DP2max 12,288 tokens/GPU, CPU optimizer offload (fraction 0.9)	#2 of 3	ATTUNE

Wanted: recipes for these models

Submit a recipe ›

QwenQwen3.8-Flash-NextQwen3.5-27BQwen3-Next-80B-A3BQwen3-235B-A22BGLMGLM-5.3-FlashGLM-5.3GLM-5.1GLM-5KimiKimi-K3Kimi-K2.5Kimi-K2DeepSeekDeepSeek-V4.1-FlashDeepSeek-V4-ProDeepSeek-V3.2DeepSeek-V3NemotronNemotron-3-Ultra-550B-A55BMiMoMiMo-V2.6-FlashInklingInkling