Best open models

Best model for each task

Each task judged on its own benchmark: the best overall, and the best that loads on a 24 GB and a 12 GB card.

Scores are Epoch AI’s own evaluations (CC BY 4.0), the best of the settings it ran for each model. Coding is measured on few, mostly large models so far.

Best model for each memory size

The most intelligent scored models that load in that much graphics memory at Q4_K_M with an 8K context.

Memory	Most intelligent that fits	Runner-up	Scored that fit
8 GB	Qwen3.5 9BECI 139.5	Llama 3.1 8B	2
12 GB	Qwen3.5 9BECI 139.5	Phi-4 14B	3
16 GB	Qwen3.5 9BECI 139.5	gpt-oss 20b	7
24 GB	Qwen3.8 27BECI 149.4	Qwen3.6 27B	15
32 GB	Qwen3.8 27BECI 149.4	Qwen3.6 27B	15
48 GB	Qwen3.8 27BECI 149.4	Qwen3.6 27B	16
64 GB	Qwen3.8 27BECI 149.4	Qwen3.6 27B	18
96 GB	Qwen3.8 27BECI 149.4	Qwen3.6 27B	19
128 GB	Qwen3.8 27BECI 149.4	Qwen3.6 27B	19
192 GB	DeepSeek V4 Flash 0731ECI 154.5	GLM 5.3 Flash	27
256 GB	DeepSeek V4 Flash 0731ECI 154.5	GLM 5.3 Flash	30
512 GB	DeepSeek V4.1 FlashECI 154.9	DeepSeek V4 Flash 0731	38

• DeepSeek V4.1 Flash436 GB#3ECI 154.9
• DeepSeek V4 Pro 0813887 GB#2ECI 155.3
• GLM 5.3 Flash176 GB#5ECI 151.9
• Qwen3.8 27B17 GB#11ECI 149.4
• Qwen3.8 Flash Next112 GBNot scored yet
• Granite 4.2 3B3 GBNot scored yet
• Granite 4.2 8B7 GBNot scored yet
• Granite 4.2 30B19 GBNot scored yet
• Nemotron 3.5 Lightning 30B A3B24 GBNot scored yet

Open vs. closed models

Both sides on the same index (Epoch Capabilities Index, CC BY 4.0). Release dates are Epoch’s.