Deploy Open Models on GPUs

Choose a GPU setup and a capability to find the best open models that fit, and deploy one as an API.

Reasoning & Math · 8× NVIDIA RTX 5090

Ranked by benchmark scores for Competition math, logical and scientific reasoning

#	Model	Params	Precision	Max context	Score (0–10)
1	MiniMax-M2.5 2026.03	228.7B	NVFP4	192K	9.93
2	Step-3.5-Flash 2026.03	199.4B	FP8	124K	9.77
3	MiniMax-M2.1 2026.02	228.7B	NVFP4	192K	9.75
4	DeepSeek-V4-Flash 2026.07	284B	NVFP4	439K	9.63
5	Qwen3.5-397B-A17B 2026.02	397B	NVFP4	70K	9.41
6	MiniMax-M2 2025.12	228.7B	NVFP4	192K	9.34
7	gpt-oss-120b 2025.08	116.8B	FP8	128K	8.90
8	DeepSeek-V4-Flash-0731 2026.08	304.2B	NVFP4	242K	8.70
9	Mistral-Medium-3.5-128B 2026.07	127.7B	FP8	256K	8.20
10	Qwen3-Coder-Next 2026.03	80B	FP16	256K	8.20
11	Qwen3.5-122B-A10B 2026.02	122B	FP8	256K	7.85
12	MiniMax-M2.7 2026.04	228.7B	NVFP4	200K	7.80
13	gpt-oss-safeguard-120b 2025.10	120.4B	FP8	128K	7.60
14	Leanstral 1.5 2026.06	119B	FP8	256K	7.50
15	Mistral Medium 3.5 2026.04	128B	FP8	256K	7.50
16	Ling 2.6 Flash 2026.04	104B	FP16	256K	7.50
17	Mistral Small 4 2026.03	119B	FP8	256K	7.50
18	Nemotron 3 Super 2026.03	120B	FP8	976K	7.50
19	Leanstral 1.0 2026.03	119B	FP8	256K	7.20
20	Qwen-AgentWorld-35B-A3B 2026.06	34.7B	FP16	256K	7.10

Deployment verified on HyperAI

Scores aggregate public benchmarks for the selected capability.
Fits usable VRAM with weights plus KV cache at 64K+ context, at the best precision.