IBM · granitemoe
Can I run granite3-moe?
The IBM Granite 1B and 3B models are the first mixture of experts (MoE) Granite models from IBM designed for low latency usage.
Parameter counts and memory belong to each build below. A family can contain several sizes and tasks; fitting in memory is not a quality ranking.
Calculate memory for a build and contextgranite3-moe system requirements
Memory needed at 4K context, counting model weights, KV cache, compute buffers and runtime overhead. Download sizes come from the official Ollama registry.
| Build | Quantization | Download | Memory needed |
|---|---|---|---|
| granite3-moe:1b-instruct-q4_01B parameters (tag)4K documented maximum contextBuild sourceCatalog checked 2026-10-13 | Q4_0 | 0.7 GB | ~1.4 GB |
| granite3-moe:1b1B parameters (tag)4K documented maximum contextBuild sourceCatalog checked 2026-10-13 | Q4_K_M | 0.8 GB | ~1.5 GB |
| granite3-moe:1b-instruct-q5_K_M1B parameters (tag)4K documented maximum contextBuild sourceCatalog checked 2026-10-13 | Q5_K_M | 0.9 GB | ~1.6 GB |
| granite3-moe:1b-instruct-q6_K1B parameters (tag)4K documented maximum contextBuild sourceCatalog checked 2026-10-13 | Q6_K | 1.0 GB | ~1.7 GB |
| granite3-moe:1b-instruct-q8_01B parameters (tag)4K documented maximum contextBuild sourceCatalog checked 2026-10-13 | Q8_0 | 1.3 GB | ~2.0 GB |
| granite3-moe:3b-instruct-q4_03B parameters (tag)4K documented maximum contextBuild sourceCatalog checked 2026-10-13 | Q4_0 | 1.8 GB | ~2.6 GB |
| granite3-moe:3b3B parameters (tag)4K documented maximum contextBuild sourceCatalog checked 2026-10-13 | Q4_K_M | 1.9 GB | ~2.7 GB |
| granite3-moe:3b-instruct-q5_K_M3B parameters (tag)4K documented maximum contextBuild sourceCatalog checked 2026-10-13 | Q5_K_M | 2.2 GB | ~3.1 GB |
Showing the 8 smallest non-alias builds. Choose any build and change the context. A context above the build's documented maximum is a hypothetical calculation, not a supported setting.
How to run granite3-moe locally
- 1.Install Ollama for Windows, macOS or Linux.
- 2.Download this exact build in a terminal:
ollama pull granite3-moe:1b-instruct-q4_0 - 3.Check the source for the build's task. For text generation, run
ollama run granite3-moe:1b-instruct-q4_0. Base/text builds may need a prompt template; an instruct build is usually the appropriate starting point for a conversation.
Frequently asked questions about granite3-moe
- How much VRAM do I need to run granite3-moe?
- At 4K context, the smallest published build of granite3-moe (Q4_0) needs roughly 1.4 GB of GPU memory once weights, KV cache, compute buffers and runtime overhead are counted. Bigger quantizations and longer context windows need more. It can also run partly or entirely on the CPU using system RAM, more slowly.
- Can I run granite3-moe without a dedicated GPU?
- A supported CPU runtime can use system RAM. The smallest listed build needs an estimated 1.4 GB at 4K context, plus headroom for your operating system. Check the build's runtime requirements. We have not measured its speed on your CPU.
- How big is the granite3-moe download?
- The smallest published build is 0.7 GB. There are 12 builds in total, the largest being 6.3 GB. Leave some extra free disk space beyond the download itself.
- How do I run granite3-moe locally?
- Install Ollama, then download a specific build with "ollama pull granite3-moe:1b-instruct-q4_0". For a build that supports text generation, use ollama run with the same tag. Check the model source for its task and prompt format.
Will granite3-moe run on your PC?
Scan your hardware once and get a verdict for this model and every build — free, no account.
Compare with similar models
Or start from your memory: 8 GB VRAM, 12 GB VRAM, 16 GB VRAM, 24 GB VRAM or a 16 GB Mac.
Model data from the official Ollama registry · verified 13/10/2026