Use this source-aware estimator before downloading a local model. It uses the exact Ollama tag, registry artifact bytes, quantization, context length and published architecture metadata. It does not inspect your computer or promise a response speed.
HiseHub LLM VRAM Calculator
Can your computer run this local AI model?
Estimate the memory pressure for one verified Ollama model, quantization, and context setting. Enter capacity, reservation and runtime buffer in GiB (1 GiB = 1,073,741,824 bytes). The catalog’s download sizes remain labelled decimal GB; the calculator converts their exact bytes to GiB. The result is a planning estimate—not a benchmark or a promise of speed.
Choose a model and enter your available capacity to see a transparent estimate.
Compare context lengths
See how 4K, 8K, 16K and 32K contexts change the planning estimate for the selected model.
| Context | KV cache | Estimated total | Remaining | Action |
|---|---|---|---|---|
| Select a verified model to compare context lengths. | ||||
What this estimate includes
- The exact Ollama artifact bytes for the tag, converted to GiB, plus the editable runtime buffer. Download size is not a promise of exact runtime VRAM.
- For supported full-attention GQA snapshots, a KV-cache estimate derived from the published architecture metadata, context length, and an FP16 cache assumption.
- A separate capacity check for GPU VRAM, system RAM, or Apple unified memory. These pools are not added together.
- Gemma 3 is listed but intentionally not estimated: its 5:1 local sliding-window/global attention and multimodal path need a hybrid cache model that this simple formula does not yet verify safely.
Important: a model that fits may still be slow, may be partly offloaded, or may require a compatible driver/backend. Images, multiple concurrent requests, long prompts, Flash Attention, KV-cache quantization, and other applications change the result. Use ollama ps after loading to verify the actual processor split.
Catalog verified September 16, 2026. Sources are linked beside each model. The tool does not detect your device, contact localhost, upload configuration, or require an API key.
| Model tag | Artifact download (GB decimal) | Quantization | Max context | Source |
|---|---|---|---|---|
llama3.2:3b-instruct-q4_K_M | 2 GB | Q4_K_M | 128K | Official Ollama page |
qwen2.5:7b-instruct-q4_K_M | 4.7 GB | Q4_K_M | 32K | Official Ollama page |
qwen2.5:14b-instruct-q4_K_M | 9 GB | Q4_K_M | 32K | Official Ollama page |
gemma3:4b-it-q4_K_M | 3.3 GB | Q4_K_M | 128K | Official Ollama page |
gemma3:12b-it-q4_K_M | 8.1 GB | Q4_K_M | 128K | Official Ollama page |
llama3.1:8b-instruct-q4_K_M | 4.9 GB | Q4_K_M | 128K | Official Ollama page |
How to interpret the result
The estimate separates the model artifact, KV cache and an editable runtime buffer. Artifact size comes from the official Ollama registry manifest and is converted from bytes to GiB. The KV estimate uses the model’s published layers, KV heads, key/value width and an FP16 assumption. A result that fits means only that the selected planning assumptions leave room in the chosen memory pool.
Context length increases KV-cache memory. Images, multiple concurrent requests, long prompts, backend support, drivers, CPU offload and other applications can change the actual run. The 4K/8K/16K/32K comparison is a planning aid, not a benchmark.
For the final check, load the same model tag locally and run ollama ps. Ollama documents that the command shows the processor split and allocated context. Read the context-length documentation and FAQ before treating a borderline result as usable.
Where this fits in HiseHub
For deployment choices, compare the self-hosted AI tools directory. If local hardware is not the right trade-off, compare API cost assumptions with subscription versus API usage. Those pages answer different questions; this calculator only estimates memory pressure from a small, verified model catalog.
Scope: this version includes six verified Ollama Q4_K_M snapshots with enough official architecture metadata for the displayed formula. It is deliberately not a full GPU compatibility database, benchmark, device detector or guarantee that a model will run at an acceptable speed. Unsupported or missing data remains unavailable rather than being treated as zero. HiseHub does not upload your configuration or require an API key.