AMD · CDNA 2

Best AI Models for AMD Instinct MI250X (128 GB)

Runs everything up to 120B MoE well.

128 GB HBM2e3277 GB/s383 TFLOPS FP16560 WCDNA 2 2021

What LLMs can the Instinct MI250X run?

#	Model	Speed	Quant · VRAM
1	Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 113 t/s	113 tok/s	Q4_K_M · 17.9 GB
2	Inkling SmallThinking Machines IQ3_S · 116.2 GB · 114 t/s	114 tok/s	IQ3_S · 116.2 GB
3	Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 113 t/s	113 tok/s	Q4_K_M · 17.9 GB
4	Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB · 96.4 t/s	96.4 tok/s	Q4_K_M · 21.7 GB
5	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 89.8 t/s	89.8 tok/s	Q4_K_M · 19.3 GB

#	Model	Speed	Quant · VRAM
1	Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 113 t/s	113 tok/s	Q4_K_M · 17.9 GB
2	Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB · 96.4 t/s	96.4 tok/s	Q4_K_M · 21.7 GB
3	Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB · 106 t/s	106 tok/s	Q4_K_M · 22.0 GB
4	Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 135 t/s	135 tok/s	Q4_K_M · 16.5 GB
5	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 89.8 t/s	89.8 tok/s	Q4_K_M · 19.3 GB

#	Model	Speed	Quant · VRAM
1	Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 113 t/s	113 tok/s	Q4_K_M · 17.9 GB
2	Inkling SmallThinking Machines IQ3_S · 116.2 GB · 114 t/s	114 tok/s	IQ3_S · 116.2 GB
3	Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB · 96.4 t/s	96.4 tok/s	Q4_K_M · 21.7 GB
4	Qwen3.5 27BAlibaba Q4_K_M · 17.9 GB · 113 t/s	113 tok/s	Q4_K_M · 17.9 GB
5	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 89.8 t/s	89.8 tok/s	Q4_K_M · 19.3 GB

#	Model	Speed	Quant · VRAM
1	Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 113 t/s	113 tok/s	Q4_K_M · 17.9 GB
2	Inkling SmallThinking Machines IQ3_S · 116.2 GB · 114 t/s	114 tok/s	IQ3_S · 116.2 GB
3	Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 113 t/s	113 tok/s	Q4_K_M · 17.9 GB
4	Qwen3.6 35B A3BAlibaba Q4_K_M · 22.0 GB · 106 t/s	106 tok/s	Q4_K_M · 22.0 GB
5	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 89.8 t/s	89.8 tok/s	Q4_K_M · 19.3 GB

#	Model	Speed	Quant · VRAM
1	Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 113 t/s	113 tok/s	Q4_K_M · 17.9 GB
2	Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 113 t/s	113 tok/s	Q4_K_M · 17.9 GB
3	Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB · 96.4 t/s	96.4 tok/s	Q4_K_M · 21.7 GB
4	Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 135 t/s	135 tok/s	Q4_K_M · 16.5 GB
5	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 89.8 t/s	89.8 tok/s	Q4_K_M · 19.3 GB

#	Model	Speed ↓	Quant · VRAM
1	Qwen3.5 4BAlibabaHow to run → Q4_K_M · 3.4 GB · 603 t/s	603 tok/s	Q4_K_M · 3.4 GB
2	Qwen3.5 9BAlibaba Q4_K_M · 6.6 GB · 309 t/s	309 tok/s	Q4_K_M · 6.6 GB
3	GPT OSS 20BOpenAI Q4_K_M · 13.1 GB · 164 t/s	164 tok/s	Q4_K_M · 13.1 GB
4	Gemma 4 26B A4B ITGoogle Q4_K_M · 16.5 GB · 135 t/s	135 tok/s	Q4_K_M · 16.5 GB
5	NVIDIA Nemotron 3.5 Lightning 30B A3B BF16NVIDIA Q4_K_M · 19.3 GB · 89.8 t/s	89.8 tok/s	Q4_K_M · 19.3 GB

#	Model	Speed	Quant · VRAM
1	Qwen3.8 27BAlibabaHow to run → Q4_K_M · 17.9 GB · 113 t/s	113 tok/s	Q4_K_M · 17.9 GB
2	Inkling SmallThinking Machines IQ3_S · 116.2 GB · 114 t/s	114 tok/s	IQ3_S · 116.2 GB
3	Qwen3.6 27BAlibaba Q4_K_M · 17.9 GB · 113 t/s	113 tok/s	Q4_K_M · 17.9 GB
4	Gemma 4 31B ITGoogle Q4_K_M · 21.7 GB · 96.4 t/s	96.4 tok/s	Q4_K_M · 21.7 GB
5	NVIDIA Nemotron 3.5 Super VL 120B A12B BF16NVIDIA Q4_K_M · 87.7 GB · 73.4 t/s	73.4 tok/s	Q4_K_M · 87.7 GB

See all 402 models on the Instinct MI250X →Best quant that fits · speeds estimated on this card

About the Instinct MI250X

The Instinct MI250X runs everything up to 120B MoE well. CDNA 2 generation, released 2021.

GenerationCDNA 2released 2021

Instinct MI250X strengths and weaknesses

Strengths

Runs 120B MoE models: GPT OSS 120B at 112 tok/s in 70.8 GB
Room for long context: Llama 3.3 70B Instruct at Q4 uses 51.3 GB, leaving 76.7 GB free

Weaknesses

560 W under load: Plan a 1,150 W PSU; ≈$123 a year at 4 h a day
ROCm, not CUDA: No FP8 weights; MXFP4 only through llama.cpp, not vLLM

Instinct MI250X tokens per second by model size

Output speed depends on memory bandwidth (3277 GB/s).

9BQwen3.5 9B · Q4_K_M · 6.6 GB
309tok/s
14BQwen3 14B · Q4_K_M · 11.9 GB
207tok/s
27–31BQwen3.8 27B · Q4_K_M · 17.9 GB
113tok/s
35B MoEQwen3.6 35B A3B · Q4_K_M · 22.0 GB
106tok/s
70BLlama 3.3 70B Instruct · Q4_K_M · 51.3 GB
42.2tok/s
120B MoEGPT OSS 120B · Q4_K_M · 70.8 GB
112tok/s

Estimated speeds: Q4_K_M, 4K context, one user. How we estimate

All 1291 models on the Instinct MI250X

Featured only

Best quant that fits, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works

LLM models compatible with AMD Instinct MI250X — ranked by Score, recency and Fit
Model	Quant	VRAM	Speed	Context	Score	Fit
Deepseek Coder 1.3B Base1.3BChatCode Q4_K_M·1328.4 t/s	Q4_K_M	4.3 GB3%	1328.4 t/s	16K		S100, Runs great

Instinct MI250X specs, power and PSU

Specs

Architecture: CDNA 2
VRAM: 128 GB HBM2e
Memory bandwidth: 3276.8 GB/s
FP16 (dense): 383.00 TFLOPS
Stream processors: 14,080
TDP: 560 W
Released: 2021-11-08

Instinct MI250X FAQ

What is the largest model the Instinct MI250X can run?

MiniMax M2.7 (228.7B) is the largest at 4-bit or better: at IQ4_XS it uses 125.3 GB of 128 GB, and runs at about 64.8 tok/s. At 3-bit, a bigger model fits too: Inkling Small (266.0B) at IQ3_S, 116.2 GB.

Can the Instinct MI250X run Qwen3.8 27B?

Yes. Qwen3.8 27B fits at Q4_K_M, using 17.9 GB of 128 GB, and it is the best model that runs well on this card.

How many tokens per second does the Instinct MI250X give?

About 113 tok/s on Qwen3.8 27B at Q4_K_M, with the first token after 2.7 s on a 4K-token prompt.

What power supply does the Instinct MI250X need?

Plan a 1150 W PSU (2× the 560 W TDP).

Is 128 GB enough for coding and agents?

Yes for coding: Qwen3.8 27B, the best of the 14 coding models that fit, runs at about 113 tok/s.

Best coding LLMs 48 GB VRAM GPUs with 128 GB or more RTX 5090 RTX 4090