Instinct MI100 32G — what LLMs can it run?
86 of 110 indexed models fit comfortably in 32GB at 4K context, each at the highest-quality quant that still leaves headroom.
The short answer for Instinct MI100 32G
32 GB, 1,228 GB/s HBM2. 86 of 110 models in this index fit comfortably at 4K context; 89 load at all.
- Biggest that fits
- Nemotron 3.5 Bolt 74B-A9B — 23.1 GB at IQ2_XXS — 9.0 GB spare at 4K, so longer context comes out of a thin margin.
- Room to grow
- GLM-4.7-Flash — 19.2 GB at Q4_K_M — under 60% of the card, which leaves 12.8 GB for a long context window or a second process.
- Speed ceiling
- Llama 3.1 8B Instruct at Q4_K_M reads 4.6 GB of weights per token, so 1,228 GB/s puts a hard ceiling near 266 tok/s. That is arithmetic on two published numbers, not a benchmark — real throughput lands below it. No run on this card has been measured here, so there is nothing to compare the ceiling against.
- The wall you will hit
- Backend coverage, not capacity. At 1,228 GB/s the memory system is competitive with NVIDIA cards of the same size; what varies is whether your runtime has a working ROCm build for your kernel and card. llama.cpp and vLLM both ship official ROCm support — this is a setup question, not a hardware ceiling.
70B+ · 1
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Nemotron 3.5 Bolt 74B-A9B74B-A9B | IQ2_XXS | 23.05 GB | +9 GB | ≤483Ceiling |
32B · 28
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Seed-OSS 36B Instruct36B | Q5_K_M | 27.81 GB | +4.2 GB | ≤51Ceiling |
| Qwen3.6 35B-A3B35B-A3B | Q4_K_M | 23.38 GB | +8.6 GB | ≤688Ceiling |
| Command R 35B35B | Q4_K_M | 22.86 GB | +9.1 GB | ≤61Ceiling |
| Yi 1.5 34B Chat34B | Q4_K_M | 22.82 GB | +9.2 GB | ≤62Ceiling |
| Qwen3 32B Instruct32B | Q4_K_M | 21.85 GB | +10.1 GB | 21–28Derived |
| Qwen2.5 32B Instruct32B | Q4_K_M | 21.69 GB | +10.3 GB | 21–28Derived |
| Qwen2.5-Coder 32B Instruct32B | Q4_K_M | 21.69 GB | +10.3 GB | 21–28Derived |
| DeepSeek-R1-Distill-Qwen-32B32B | Q4_K_M | 21.69 GB | +10.3 GB | 21–28Derived |
| Nemotron 3.5 Lightning 30B-A3B30B-A3B | Q4_0 | 19.79 GB | +12.2 GB | ≤720Ceiling |
| Gemma 4 31B IT31B | Q4_K_M | 21.08 GB | +10.9 GB | 22–30Derived |
| Qwen3 30B-A3B Instruct30B-A3B | Q5_K_M | 23.04 GB | +9 GB | ≤607Ceiling |
| Qwen3-Coder 30B-A3B Instruct30B-A3B | Q5_K_M | 23.04 GB | +9 GB | ≤607Ceiling |
| Qwen3-VL 30B-A3B Instruct30B-A3B | Q4_K_M | 19.73 GB | +12.3 GB | ≤711Ceiling |
| GLM-4.7-Flash30B-A3B | Q4_K_M | 19.19 GB | +12.8 GB | ≤711Ceiling |
| Granite 4.2 30B30B | Q6_K | 26.19 GB | +5.8 GB | 17–23Derived |
| Muse Glimmer 30B30B | Q6_K | 24.61 GB | +7.4 GB | 18–24Derived |
| Qwen3.8 27B27B | Q4_K_M | 17.47 GB | +14.5 GB | 25–34Derived |
| Qwen3.6 27B27B | Q5_K_M | 22.19 GB | +9.8 GB | 20–26Derived |
| Gemma 3 27B IT27B | Q4_K_M | 18.37 GB | +13.6 GB | 25–33Derived |
| Gemma 2 27B Instruct27B | Q5_K_M | 21.76 GB | +10.2 GB | 22–29Derived |
| Gemma 4 26B-A4B IT26B-A4B | Q4_K_M | 16.39 GB | +15.6 GB | ≤533Ceiling |
| Mistral Small 24B Instruct24B | Q4_K_M | 15.89 GB | +16.1 GB | 29–38Derived |
| LFM2.5 24B-A2B24B-A2B | Q8_0 | 26.73 GB | +5.3 GB | ≤496Ceiling |
| Devstral Small 1.1 24B24B | Q6_K | 20.91 GB | +11.1 GB | 22–29Derived |
| Magistral Small 1.2 24B24B | Q6_K | 20.91 GB | +11.1 GB | 22–29Derived |
| Codestral 22B22B | Q4_K_M | 15.03 GB | +17 GB | 31–41Derived |
| ERNIE 4.5 21B-A3B21B-A3B | Q8_0 | 24.44 GB | +7.6 GB | ≤406Ceiling |
| GPT-OSS 20B21B MoE | MXFP4 | 12.76 GB | +19.2 GB | ≤620Ceiling |
14B · 18
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| InternLM2 20B Chat20B | Q5_K_M | 15.59 GB | +16.4 GB | 30–39Derived |
| DeepSeek-Coder-V2-Lite Instruct16B-A2.4B | Q8_0 | 17.56 GB | +14.4 GB | ≤507Ceiling |
| DeepSeek-V2-Lite Chat16B-A2.4B | Q4_K_M | 10.08 GB | +21.9 GB | ≤889Ceiling |
| StarCoder2 15B15B | Q4_K_M | 10.19 GB | +21.8 GB | 46–61Derived |
| Qwen3 14B Instruct14B | Q5_K_M | 11.65 GB | +20.4 GB | 40–53Derived |
| Qwen2.5 14B Instruct14B | Q5_K_M | 11.73 GB | +20.3 GB | 40–53Derived |
| DeepSeek-R1-Distill-Qwen-14B14B | Q4_K_M | 10.14 GB | +21.9 GB | 47–62Derived |
| Phi-4 14B14B | Q5_K_M | 11.77 GB | +20.2 GB | 40–53Derived |
| Phi-3 Medium 14B Instruct14B | Q6_K | 12.86 GB | +19.1 GB | 36–48Derived |
| Mistral Nemo 12B Instruct12B | Q6_K | 11.14 GB | +20.9 GB | 42–55Derived |
| Gemma 3 12B IT12B | Q5_K_M | 9.84 GB | +22.2 GB | 48–64Derived |
| Mellum2.1 12B-A2.5B Thinking12B-A2.5B | Q8_0 | 13.6 GB | +18.4 GB | ≤487Ceiling |
| Stable LM 2 12B Chat12B | Q4_K_M | 8.35 GB | +23.7 GB | 57–75Derived |
| Llama 3.2 11B Vision Instruct11B | Q8_0 | 12.9 GB | +19.1 GB | 36–47Derived |
| Solar 10.7B Instruct11B | Q4_K_M | 7.6 GB | +24.4 GB | 65–85Derived |
| Falcon 3 10B Instruct10B | Q4_K_M | 7.28 GB | +24.7 GB | 66–88Derived |
| Gemma 2 9B Instruct9B | Q8_0 | 11.7 GB | +20.3 GB | 43–56Derived |
| Qwen3.5 9B9B | Q8_0 | 10.37 GB | +21.6 GB | 43–56Derived |
7B · 25
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| GLM-4-9B-Chat9B | Q8_0 | 10.16 GB | +21.8 GB | 44–58Derived |
| Qwen3-VL 8B Instruct8B | Q8_0 | 10.39 GB | +21.6 GB | 45–59Derived |
| Granite 4.2 8B8B | Q8_0 | 10.45 GB | +21.6 GB | 45–59Derived |
| Ministral 3 8B Instruct8B | Q8_0 | 10.02 GB | +22 GB | 46–61Derived |
| Qwen2-VL 7B Instruct7B | Q4_K_M | 5.49 GB | +26.5 GB | 83–110Derived |
| Granite 3.1 8B Instruct8B | Q4_K_M | 5.88 GB | +26.1 GB | 84–111Derived |
| Qwen3 8B Instruct8B | Q6_K | 7.64 GB | +24.4 GB | 62–82Derived |
| Llama 3.1 8B Instruct8B | Q8_0 | 9.47 GB | +22.5 GB | 49–65Derived |
| Nous Hermes 3 Llama 3.1 8B8B | Q4_K_M | 5.64 GB | +26.4 GB | 86–114Derived |
| Aya 23 8B8B | Q4_K_M | 5.64 GB | +26.4 GB | 86–114Derived |
| OpenChat 3.6 8B8B | Q4_K_M | 5.64 GB | +26.4 GB | 86–114Derived |
| DeepSeek-R1-Distill-Llama-8B8B | Q5_K_M | 6.51 GB | +25.5 GB | 73–97Derived |
| InternLM2 7B Chat7B | Q4_K_M | 5.45 GB | +26.6 GB | 89–118Derived |
| Qwen2.5 7B Instruct7B | Q6_K | 6.77 GB | +25.2 GB | 67–89Derived |
| Qwen2.5-Coder 7B Instruct7B | Q4_K_M | 5.07 GB | +26.9 GB | 91–120Derived |
| DeepSeek-R1-Distill-Qwen-7B7B | Q4_K_M | 5.07 GB | +26.9 GB | 91–120Derived |
| Gemma 4 E4B ITE4B | Q8_0 | 8.46 GB | +23.5 GB | 52–69Derived |
| OLMo 2 7B Instruct7B | Q8_0 | 10.3 GB | +21.7 GB | 54–71Derived |
| OLMo 3 7B Instruct7B | Q8_0 | 10.3 GB | +21.7 GB | 54–71Derived |
| WizardLM-2 7B7B | Q4_K_M | 5.14 GB | +26.9 GB | 95–126Derived |
| Mistral 7B Instruct v0.37B | Q6_K | 6.75 GB | +25.3 GB | 70–93Derived |
| Zephyr 7B Beta7B | Q6_K | 6.75 GB | +25.3 GB | 70–93Derived |
| Gemma 4 E2B ITE2B | Q8_0 | 5.2 GB | +26.8 GB | ≤262Ceiling |
| Qwen3.5 4B4B | Q8_0 | 4.97 GB | +27 GB | ≤280Ceiling |
| Gemma 3 4B IT4B | Q8_0 | 5.05 GB | +27 GB | ≤283Ceiling |
≤3B · 14
| Model | Quant | Est. VRAM | Headroom | tok/s |
|---|---|---|---|---|
| Qwen3 4B Instruct4B | Q6_K | 4.05 GB | +28 GB | ≤394Ceiling |
| Phi-4 Mini Instruct3.8B | Q8_0 | 4.81 GB | +27.2 GB | ≤316Ceiling |
| Phi-3.5 Mini Instruct3.8B | Q8_0 | 5.89 GB | +26.1 GB | ≤318Ceiling |
| Granite 4.2 3B3B | Q8_0 | 4.41 GB | +27.6 GB | ≤333Ceiling |
| Llama 3.2 3B Instruct3B | Q8_0 | 4.05 GB | +28 GB | ≤379Ceiling |
| Qwen2.5 3B Instruct3B | Q8_0 | 3.59 GB | +28.4 GB | ≤394Ceiling |
| SmolLM3 3B3B | Q8_0 | 3.73 GB | +28.3 GB | ≤395Ceiling |
| Gemma 2 2B Instruct2B | Q8_0 | 3.34 GB | +28.7 GB | ≤467Ceiling |
| Qwen3.5 2B2B | Q8_0 | 2.23 GB | +29.8 GB | ≤620Ceiling |
| Qwen3 1.7B Instruct1.7B | Q8_0 | 2.39 GB | +29.6 GB | ≤706Ceiling |
| Qwen2.5 1.5B Instruct1.5B | Q8_0 | 1.83 GB | +30.2 GB | ≤792Ceiling |
| Llama 3.2 1B Instruct1B | Q8_0 | 1.51 GB | +30.5 GB | ≤982Ceiling |
| Qwen3.5 0.8B0.8B | Q8_0 | 0.94 GB | +31.1 GB | ≤1516Ceiling |
| Qwen2.5 0.5B Instruct0.5B | Q8_0 | 0.6 GB | +31.4 GB | ≤2506Ceiling |
How this list is built
Each row is the lowest-perplexity-loss quant of that model whose estimated total — weights plus KV cache at 4K context plus activation buffer — uses at most 88% of the card. That is the calculator's "green" threshold, so every row here has real headroom rather than only just fitting. Raise the context length and the list shortens; the calculator lets you check any combination directly. The speed column is this card: a measured run, a range scaled from this card's own Llama 2 7B benchmark, or a bandwidth ceiling. A figure that only exists for the RTX 4090 is labelled as that card.
A further 3 models load but with no headroom to spare (up to 105% of VRAM) — the Quant Hub’s GPU chips count those too, which is why its number is higher.
Measured on this card
No benchmark runs in this index were recorded on a Instinct MI100 32G. Every figure on this page is calculated from the model architecture and the quant level — treat them as estimates, not measurements.
Cards with the same budget
What fits is decided by memory, so every 32GB card of this type returns the same list. These pages are not different answers — they differ in throughput, which this index does not measure per card.
The same 32GB in system memory is a different machine — everything on this page fits there, at a speed set by your DIMMs rather than a graphics bus: 32 GB RAM (CPU)
Same list, different speed — play one sentence on both cards: Instinct MI100 32G vs Radeon AI PRO R9700 32G (Qwen3 32B Instruct)
Setup guides for this card
Step-by-step guides written for the Instinct MI100 32G's runtime or memory size, with the commands and the checks that prove the model is running on the card.
Stepping up
A Radeon PRO W7900 48G (48GB) fits 8 more of the indexed models than this card. Radeon PRO W7900 48G →
Before buying, hear the speed difference on a model both cards hold: Instinct MI100 32G vs Radeon PRO W7900 48G (Qwen3 32B Instruct) →
Common questions
What is the best local LLM for an Instinct MI100 32G?
For everyday use, GLM-4.7-Flash at Q4_K_M — about 19.2 GB of the card's 32 GB at 4K context, leaving 12.8 GB for a longer window. If you want the largest thing that will load, that is Nemotron 3.5 Bolt 74B-A9B at IQ2_XXS (23.1 GB, 9.0 GB spare). "Best" here means best fit for the memory budget — this index does not run task benchmarks, so it cannot tell you which model is smarter.
Can an Instinct MI100 32G run Llama 3.2 90B Vision Instruct?
No. Its smallest build here, Q3_K_M, needs about 46.2 GB and this card has 32.0 GB usable — short by 14.2 GB before any context beyond 4K. The largest model this card does clear is Nemotron 3.5 Bolt 74B-A9B.
How many tokens per second does an Instinct MI100 32G do on an 8B model at Q4?
This index has no measured run on this card, so it does not publish a figure. What can be stated from specifications: generating a token requires reading every weight once, Llama 3.1 8B Instruct at Q4_K_M is 4.6 GB of weights, and this card moves 1,228 GB/s — a ceiling near 266 tok/s. Batch size, context length, the runtime and how much of the model sits in cache all take you below it.
Instinct MI100 32G or RTX 5090 for local LLMs?
They do not hold the same set: Instinct MI100 32G fits 86 and RTX 5090 88 of 110 at 4K — the cards load different formats. Speed is the other difference — 1,228 GB/s against 1,792 GB/s, a 1.5× gap in how fast the weights can be read, which is what token generation is bound by. The RTX 5090 generates faster on any model both can hold.
86 of 110 indexed models fit comfortably in 32GB at 4K context, each at the highest-quality quant that still leaves headroom.
This page's figures change when the model or the runtime does.Last updated 2026-10-18 RSS → /feed.xml