Instinct MI210: 64GB VRAM — What Local LLMs Can It Run?
Check prices
The Instinct MI210 has 64GB of HBM2e memory with 1,638 GB/s of bandwidth. No used-price data yet. At Q4 quantization it runs models up to 70B parameters at roughly 20 tokens/s.
Full Instinct MI210 specifications
What LLMs can the Instinct MI210 run?
One reference model per family (its main version)| Model | Q4_K_M | Q8_0 | FP16 | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Fits | Speed t/s | Context | Fits | Speed t/s | Context | Fits | Speed t/s | Context | |
| DeepSeekDeepSeek R1 Qwen3 8B | ✓ | 139.2 | 128K | ✓ | 82.8 | 128K | ✓ | 46.6 | 128K |
| GemmaGemma 4 31B | ✓ | 36.9 | 32K | ✓ | 21.8 | 32K | ✗ | — | — |
| GLMGLM 4.7 Flash | ✓ | 157.7 | 128K | ✓ | 122.9 | 128K | ✓ | 86.4 | 32K |
| gpt-ossgpt-oss 20B | ✓ | 148.5 | 128K | ✓ | 112.5 | 128K | ✓ | 76.8 | 128K |
| KimiKimi Linear 48B A3B | ✓ | 163.3 | 128K | ✓ | 126.3 | 128K | Offload | 14.6 | — |
| LlamaLlama 3.1 8B | ✓ | 142.9 | 128K | ✓ | 84.7 | 128K | ✓ | 47.6 | 128K |
| MistralMistral 7B v0.3 | ✓ | 154.9 | 32K | ✓ | 92.7 | 32K | ✓ | 52.3 | 32K |
| QwenQwen3 8B | ✓ | 139.2 | 32K | ✓ | 82.8 | 32K | ✓ | 46.6 | 32K |
Fits when weights + KV cache + runtime overhead is at or below usable memory, single card. "Offload" on MoE models means it runs with expert weights in system RAM (only attention layers and KV cache stay in VRAM); that speed assumes 70 GB/s RAM bandwidth. The context column is the largest tier that fits.
Instinct MI210 used price history
Not enough history to draw a chart yet — price tracking just started.
Prices by platform
Buy links may carry affiliate tags; they do not change the price you pay.
FAQ
How much VRAM does the Instinct MI210 have?
The Instinct MI210 has 64GB of HBM2e memory with 1,638 GB/s of bandwidth.
Can the Instinct MI210 run a 70B model?
Yes. At Q4_K_M with an 8K context a 70B dense model needs about 43GB, which fits in the Instinct MI210's 64GB.
How much does a used Instinct MI210 cost?
No used-price data yet; it will appear here once the feed is live.