← Maviter Utility LabUtility 05 / Local model memory

Estimate which model class your computer can hold.

Choose a pinned open-weight model, quantization and context target. Maviter estimates weight, KV-cache and runtime memory without downloading a model or profiling your device.

Outcome: A memory-fit candidate, visible arithmetic and the runtime checks still required
Time: About 2 minutes
Privacy: Local by default; optional private save

Open calculator →

✓ No benchmark or speed claim✓ No model download✓ Capacity does not prove runtime compatibility

Local by default

You choose what to keep.

Calculations run locally without provider requests or input analytics. Refreshing clears working values. Only “Confirm & save snapshot” sends the displayed numbers/choices and your title to private, encrypted account storage. Nothing is automatically saved.

Catalogue revision 2026-09-01.1

Describe the memory envelope.

Select the exact pinned model record, then enter the memory available on the computer you intend to test.

1. Which exact model?2. Weight quantization

3. Requested context tokensPinned model limit: 40,960 tokens.4. System RAM (GiB)Installed memory; the estimate reserves 30% for the operating system and other work.5. GPU VRAM (GiB)Enter 0 for a CPU-only planning case; 15% is reserved from stated VRAM.

Memory-fit verdict

Full-GPU memory candidate

The estimated working set fits inside the calculator's usable-VRAM allowance.

Quantized weights: 1.87 GiB
KV cache estimate: 1.13 GiB
Runtime reserve: 1.25 GiB
Working set: 4.25 GiB
Usable VRAM: 6.8 GiB
Usable RAM: 22.4 GiB

Arithmetic

Weights = parameters × quantization bits ÷ 8. KV cache = 2 × layers × KV heads × head dimension × requested tokens × 2-byte cache. Runtime reserve = greater of 1.25 GiB or 15% of weights.

Keep this decision

Saving is optional. Calculations are free; signed-in users can keep a private snapshot. Premium and Team can compare snapshots and export briefs.

My saved calculations →

Versioned source register

Ten exact models—not a timeless “best LLM” list.

This small catalogue grounds the calculator. Every row names an exact first-party repository and pinned revision, while the separate source monitor sends publisher drift to human review. Open the complete model evidence briefs →

Catalogue revision: 2026-09-01.1
Reviewed: 1 September 2026
Review by: 15 September 2026
Selection rule: Representative memory classes, not ranking

Model metadata used by Local LLM Fit Calculator revision 2026-09-01.1
Exact model	Parameters	Documented context	Access and licence	Memory evidence	Pinned source
Gemma 3 1B ITgoogle/gemma-3-1b-itA compact instruction-tuned model for constrained local experiments.	1.0B	32,768 tokens	Gated filesGemma terms	Weight floor onlyKV result withheld	Official repository ↗`dcc83ea841ab`
Phi-4 Mini Instructmicrosoft/Phi-4-mini-instructA small instruction model with a long documented context window.	3.84B	1,31,072 tokens	Open filesMIT	Formula-backedConfig geometry recorded	Official repository ↗`cfbefacb9925`
Gemma 3 4B ITgoogle/gemma-3-4b-itA compact multimodal instruction model with gated model files.	4.30B	1,31,072 tokens	Gated filesGemma terms	Weight floor onlyKV result withheld	Official repository ↗`093f9f388b31`
Qwen3 4BQwen/Qwen3-4BA compact general instruction and reasoning model for local evaluation.	4.02B	40,960 tokens	Open filesApache-2.0	Formula-backedConfig geometry recorded	Official repository ↗`1cfa9a720891`
DeepSeek R1 Distill Qwen 7Bdeepseek-ai/DeepSeek-R1-Distill-Qwen-7BA distilled reasoning-oriented model in the familiar 7B local class.	7.62B	1,31,072 tokens	Open filesMIT	Formula-backedConfig geometry recorded	Official repository ↗`916b56a44061`
Qwen3.5 9BQwen/Qwen3.5-9BA current mid-sized Qwen model with a large native context limit.	9.65B	2,62,144 tokens	Open filesApache-2.0	Formula-backedConfig geometry recorded	Official repository ↗`c20223623576`
DeepSeek R1 Distill Qwen 14Bdeepseek-ai/DeepSeek-R1-Distill-Qwen-14BA larger distilled reasoning model for machines with more memory headroom.	14.77B	1,31,072 tokens	Open filesMIT	Formula-backedConfig geometry recorded	Official repository ↗`1df8507178af`
Mistral Small 3.2 24Bmistralai/Mistral-Small-3.2-24B-Instruct-2506A 24B instruction model intended for capable local or self-hosted systems.	24.01B	1,31,072 tokens	Open filesApache-2.0	Formula-backedConfig geometry recorded	Official repository ↗`95a6d26c4bfb`
Qwen3.6 27BQwen/Qwen3.6-27BA current dense 27B model for higher-memory local evaluation.	27.78B	2,62,144 tokens	Open filesApache-2.0	Formula-backedConfig geometry recorded	Official repository ↗`6a9e13bd6fc8`
Gemma 4 31B ITgoogle/gemma-4-31B-itA large current Gemma instruction model with substantial memory requirements.	31.27B	2,62,144 tokens	Open filesApache-2.0	Formula-backedConfig geometry recorded	Official repository ↗`842da3794eaa`

How to use the result

Estimate first. Verify the changing facts next.

The utility makes its assumptions and arithmetic visible. Before spending money or changing a live system, replace example values with current first-party information and preserve a rollback path.

01Enter the exact jobUse real workload values, not an aspirational maximum.
02Read the boundarySee what the calculator deliberately does not infer.
03Verify at sourceConfirm current limits, prices, records or plan terms.
04Make a controlled decisionRecord the owner, evidence and rollback before action.