GPUs for on-prem AI / AMD
AMD Instinct MI210 64GB for LLMs
CDNA 2 PCIe card with 64GB; no FP8, and engine and kernel coverage on ROCm lags newer Instinct parts. Price is an indicative 2026 street range for one unit, not a quote.
PCIe | Infinity Fabric bridge 3 links x 100 GB/s | ROCm | indicative price $4,000-8,000
Which models fit, and how many cards
Planning estimates. 4-bit column: one user at 8K context. FP8 column: 10 users at once at 16K context.
| Model | Size | 4-bit: cards | 4-bit: tok/s | FP8, 10 users: cards | FP8: tok/s per user |
|---|---|---|---|---|---|
| Qwen3-Coder-Next | 79.7B / 3B | 1 | 256 | 2 | 179 |
| Llama 3.3 70B Instruct | 70B | 1 | 23 | 2 | 19 |
| Qwen3.6-35B-A3B | 36B / 3B | 1 | 257 | 1 | 146 |
| Qwen3-32B | 32.8B | 1 | 47 | 1 | 21 |
| Olmo 3.1 32B Think | 32B | 1 | 48 | 1 | 24 |
| GLM-4.7-Flash | 31.2B / 3B | 1 | 250 | 1 | 113 |
| Gemma 4 31B IT | 30.7B | 1 | 49 | 1 | 22 |
| Qwen3-30B-A3B-Instruct-2507 | 30.5B / 3.3B | 1 | 229 | 1 | 84 |
| Granite 4.2 30B | 30B | 1 | 51 | 1 | 23 |
| NVIDIA Nemotron 3.5 Lightning 30B-A3B | 30B / 3B | 1 | 260 | 1 | 166 |
| Qwen3.8-27B | 27.8B | 1 | 56 | 1 | 30 |
| Gemma 3 27B IT | 27B | 1 | 57 | 1 | 28 |
| Gemma 4 26B A4B IT | 25.2B / 3.8B | 1 | 212 | 1 | 94 |
| Mistral Small 3.2 24B Instruct 2506 | 24B | 1 | 63 | 1 | 30 |
| Magistral Small 1.2 (2509) | 24B | 1 | 63 | 1 | 30 |
| Devstral Small 2 24B Instruct 2512 | 24B | 1 | 63 | 1 | 30 |
| gpt-oss-20b | 21B / 3.6B | 1 | 229 | 1 | 126 |
| Phi-4 | 14B | 1 | 99 | 1 | 40 |
| Gemma 4 12B IT | 11.95B | 1 | 114 | 1 | 54 |
| Qwen3.5-9B | 9.7B | 1 | 138 | 1 | 77 |
| Qwen3-VL-8B-Instruct | 8.8B | 1 | 144 | 1 | 58 |
| Granite 4.2 8B | 8B | 1 | 153 | 1 | 59 |
| Llama 3.1 8B Instruct | 8B | 1 | 155 | 1 | 64 |
| Step-3.7-Flash | 198B / 11B | 2 | 138 | 4 | 114 |
| Qwen3.8-Flash-Next | 180B / 6B | 2 | 196 | 4 | 159 |
| Mistral Medium 3.5 128B | 128B | 2 | 22 | 4 | 20 |
| Qwen3.5-122B-A10B | 125B / 10B | 2 | 147 | 4 | 126 |
| Ling-3.0-flash | 124B / 5.1B | 2 | 219 | 4 | 185 |
| NVIDIA Nemotron 3 Super 120B-A12B | 120B / 12B | 2 | 118 | 4 | 99 |
| Mistral Small 4 119B A6B (2603) | 119B / 6.5B | 2 | 197 | 4 | 167 |
| gpt-oss-120b | 117B / 5.1B | 2 | 223 | 4 | 176 |
| Command A (03-2025) | 111B | 2 | 26 | 4 | 25 |
| Llama 4 Scout 17B-16E Instruct | 109B / 17B | 2 | 99 | 4 | 73 |
| Trinity Large Thinking | 398B / 13B | 4 | 141 | 8 | 107 |
| GLM-5.3-Flash | 320B / 18B | 4 | 135 | 8 | 122 |
| MiMo-V2.6-Flash-RL | 309B / 15B | 4 | 145 | 8 | 127 |
| Hy3 | 295B / 21B | 4 | 102 | 8 | 72 |
| Command A+ (05-2026) | 218B / 25B | 4 | 116 | 4 | 70 |
| DeepSeek-V4.1-Flash-1017 | 748B / 16B | 8 | 171 | 16 (multi-node) | 112 |
| GLM-5.3 | 744B / 40B | 8 | 90 | 16 (multi-node) | 57 |
| Mistral Large 3 675B Instruct 2512 | 675B / 41B | 8 | 98 | 16 (multi-node) | 65 |
| DeepSeek-V3.2 | 671B / 37B | 8 | 103 | 16 (multi-node) | 68 |
| MiniMax-M3 | 428B / 23B | 8 | 126 | 8 | 89 |
| Qwen3.5-397B-A17B | 403B / 17B | 8 | 139 | 8 | 110 |
| Llama 4 Maverick 17B-128E Instruct | 400B / 17B | 8 | 153 | 8 | 102 |
| DeepSeek-V4-Pro-0813 | 1600B / 49B | 16 (multi-node) | 76 | 32 (multi-node) | 70 |
| MiMo-V2.6-Pro-RL | 1020B / 42B | 16 (multi-node) | 73 | 24 (multi-node) | 63 |
| Kimi-K2.6 | 1000B / 32B | 16 (multi-node) | 84 | 24 (multi-node) | 73 |
| Kimi-K2.7-Code | 1000B / 32B | 16 (multi-node) | 84 | 24 (multi-node) | 73 |
| Muse Spark Open 816B-A42B | 816B / 42B | 16 (multi-node) | 78 | 16 (multi-node) | 63 |
| Kimi-K3 | 2800B / 104B | 32 (multi-node) | 52 | 56 (multi-node) | 47 |
| Qwen3.8-2.4T-A95B | 2446B / 95B | 32 (multi-node) | 53 | 48 (multi-node) | 46 |
Design a deployment on AMD Instinct MI210 64GB
Preloaded with this hardware. Add your ERP, documents and users.
Questions
What is the largest model that fits on one AMD Instinct MI210 64GB?
Of the models we track, Qwen3-Coder-Next (79.7B total, 3B active) fits on a single 64 GB card in 4-bit with one user at 8K context. More users or longer context need memory for the KV cache, so leave headroom.
How fast is AMD Instinct MI210 64GB for LLM inference?
Generation speed is mostly set by memory bandwidth: 1600 GB/s here. Each token reads the active weights once, so a model with 8 GB of active weights tops out near 200 tokens per second per stream before efficiency losses; batching many users raises total throughput.
Is AMD Instinct MI210 64GB a good choice for on-prem AI?
CDNA 2 PCIe card with 64GB; no FP8, and engine and kernel coverage on ROCm lags newer Instinct parts. Price is an indicative 2026 street range for one unit, not a quote.