Models

Which model do you want to run?

Pick a model and see the bare minimum to run it, every GPU configuration that fits, AMD against NVIDIA, and what a fleet of a thousand costs, all from the model's own config and live rental prices.

21 models, 73 datacenter GPUs, refreshed with every release.

Popular models

All models

GPU counts are the tensor-parallel width of H100 SXM5 80GB parts needed at native precision: minimum is 4,096 tokens of context and one stream, recommended is 32,768 tokens and 32 streams.

Model	Lab	Parameters	Active	Context	Attention	Native	Min H100s	Recommended
GLM-5.3	Z.ai	753B	41.7B	1,048,576	MLA	FP8	16	16
GLM-5.2	Z.ai	753B	41.7B	1,048,576	MLA	BF16	Over 16	Over 16
DeepSeek-V4-Pro Estimated	DeepSeek	1.60T	49.8B	1,048,576	MLA	INT4 / FP4	16	16
DeepSeek-V4-Flash Estimated	DeepSeek	291B	13.8B	1,048,576	MLA	INT4 / FP4	4	4
Gemma 4 31B	Google	31.3B	dense	262,144	GQA + sliding	BF16	1	2
Gemma 4 26B-A4B	Google	25.8B	4.0B	262,144	GQA + sliding	BF16	1	1
Qwen3.5-122B-A10B	Alibaba	125B	10.0B	262,144	Hybrid	BF16	4	4
Qwen3.5-27B	Alibaba	27.8B	dense	262,144	Hybrid	BF16	1	2
Qwen3.5-397B-A17B	Alibaba	403B	17.0B	262,144	Hybrid	BF16	16	16
gpt-oss-120b	OpenAI	117B	5.1B	131,072	GQA + sliding	INT4 / FP4	1	2
gpt-oss-20b	OpenAI	20.9B	3.6B	131,072	GQA + sliding	INT4 / FP4	1	1
GLM-4.5	Z.ai	358B	32.0B	131,072	GQA	BF16	16	16
Kimi K2 Instruct	Moonshot AI	1.03T	32.0B	131,072	MLA	FP8	16	16
Mistral Small 3.2 24B	Mistral AI	24.0B	dense	131,072	GQA	BF16	1	4
Qwen3-235B-A22B	Alibaba	235B	22.0B	40,960	GQA	BF16	8	16
Qwen3-32B	Alibaba	32.8B	dense	40,960	GQA	BF16	1	8
Gemma 3 27B	Google	27.4B	dense	131,072	GQA + sliding	BF16	1	4
DeepSeek-R1	DeepSeek	684B	37.0B	163,840	MLA	FP8	16	16
Phi-4	Microsoft	14.7B	dense	16,384	GQA	BF16	1	2
Llama 3.3 70B Instruct	Meta	70.6B	dense	131,072	GQA	BF16	2	8
Llama 3.1 8B Instruct	Meta	8.0B	dense	131,072	GQA	BF16	1	4

Which model do you want to run?

Popular models

All models

Stay Updated on GPU Specs