FreeLLMAll models › ibm-granite
granite-3.2-8b-instruct
ibm-granite · open license · released 17 Feb 2025
Download No Arena votes
Arena score—Not voted on Arena yet
Size8.2 Bparameters
4-bit download5.0 GBweights only, Q4_K_M
Max context128Ktokens, from the model config
Where to use it for free
- Run it in one command with whichllm (free, open source):
uvx whichllm@latest plan "granite-3.2-8b-instruct"shows the GPU you need,uvx whichllm@latest run "granite-3.2-8b-instruct gguf"downloads it and starts a chat - Download and run it at home: official weights · GGUF files for Ollama and LM Studio
No Arena votes yet
Arena has not voted on this model. It is listed because it is a popular small model on its lab's official Hugging Face page (51K downloads a month).
Speed on popular hardware
| Hardware | GPU memory | Fits? | Speed |
|---|---|---|---|
| RTX 3060 12 GB | 12 GB | Fits in GPU | 26 tok/s |
| RTX 4060 Ti 16 GB | 16 GB | Fits in GPU | 21 tok/s |
| RTX 3090 | 24 GB | Fits in GPU | 66 tok/s |
| RTX 4090 | 24 GB | Fits in GPU | 71 tok/s |
| RTX 5090 | 32 GB | Fits in GPU | 119 tok/s |
| RX 7900 XTX | 24 GB | Fits in GPU | 68 tok/s |
| 2 × RTX 3090 | 48 GB | Fits in GPU | 64 tok/s |
| 4 × RTX 3090 | 96 GB | Fits in GPU | 61 tok/s |
| Mac M4 Max 128 GB | 96 GB | Fits in GPU | 36 tok/s |
| Ryzen AI Max+ 395 128 GB | 96 GB | Fits in GPU | 19 tok/s |
| 8 × H100 SXM | 640 GB | Fits in GPU | 167 tok/s |
Estimate at 4-bit with 32K context, 64 GB of DDR5 for offload. Speed ≈ memory bandwidth × 78% ÷ data read per token, plus a fixed cost per token, calibrated on 44 real llama.cpp tests; typical error about ±10%. Compare all GPUs.
GPU memory needed
| Level | 8K context | 32K context | 128K context | Quality vs full |
|---|---|---|---|---|
| Q3_K_M | 6.0 GB | 10 GB | 26 GB | +10.4% noticeable loss |
| Q4_K_M | 6.9 GB | 11 GB | 27 GB | +2.8% small loss |
| Q5_K_M | 7.8 GB | 12 GB | 28 GB | +0.9% very small loss |
| Q6_K | 8.7 GB | 13 GB | 29 GB | +0.3% near-lossless |
| Q8_0 | 11 GB | 15 GB | 31 GB | ≈ 0% same as full |
| F16 | 18 GB | 22 GB | 38 GB | ≈ 0% same as full |
Weights + context cache (from the model's architecture) + runtime. Quality: increase in perplexity against the full model, from llama.cpp tests, scaled by model size. Simulate your own setup.
Share this page:XLinkedInRedditWhatsApp
Spotted a wrong number or a broken link? Report a mistake