Methodology
Every estimate has a visible edge.
LocalLM Fit estimates whether a model is likely to fit. It does not pretend to know the exact peak memory or tokens per second of every runtime on every machine.
The memory formula
required memory = published weights + KV cache + runtime overhead KV cache = kvCacheGbPer8k × (selected context ÷ 8,192) runtime overhead = 1.5 GB The calculator compares that requirement with the usable accelerator memory in the selected hardware profile. On Apple silicon, the accelerator figure is unified memory. On a discrete GPU, it is VRAM rather than system RAM.
The three result states
| Result | Available ÷ required | Interpretation |
|---|---|---|
| Comfortable | ≥ 1.2× | At least 20% headroom after the estimate. This is the preferred target for a new purchase. |
| Tight | ≥ 1.0× and < 1.2× | The estimate fits, but there is limited room for the operating system, display, other applications, or longer contexts. |
| Won't fit | < 1.0× | The selected memory is below the estimate. A smaller model, lower-bit quantization, or more memory is required. |
MoE models can receive one additional state: Tight fit · CPU offload. This occurs when the full weights do not fit in accelerator memory but do fit in system memory. The result remains tight because CPU or mixed CPU/GPU inference is expected to be substantially slower.
How KV cache is estimated
For most models, the per-8K figure is a conservative FP16 estimate derived from the public attention configuration: layer count, KV heads, head dimension, and the model's attention pattern. Sliding-window or hybrid-attention models can use less than the displayed estimate. DeepSeek V4.1 Flash uses the published 890 bytes/token figure where available.
KV cache grows with the selected context. It is not a fixed model-file cost. Runtime quantization, paged attention, prefix caching, concurrency, and backend-specific workspace can change the actual peak.
Quantization and hardware targets
Weight sizes come from published GGUF file listings rather than parameter-count multiplication. For models with separate multimodal projectors, the listed size includes the published projector file where available. The VRAM guides use an 8K context and choose the highest-ranked published quantization that still lands in the comfortable or tight band for that model. The preference order is Q8_0, Q5_K_M, Q4_K_M, Q4_0, MXFP4, then the published low-bit IQ tiers.
Hardware recommendations use editorial price classes, not live prices. A target is selected only when its memory is at least 1.2× the calculated requirement. Product availability, regional pricing, and exact model configuration must be checked at the retailer.
Accessory recommendations
Accessory blocks use broad product categories and Amazon search links. They do not publish invented prices, ratings, availability, or compatibility guarantees. Product pages are linked through the site's affiliate mechanism only after the category and use case are stated.
Storage, docks, cooling, power protection, and desktop memory can improve a local AI setup, but none increases GPU VRAM or Apple unified memory. An accessory is never presented as a way to make an oversized model fit.
Model sources and verification dates
Hardware sources and verification dates
Update log
- LocalLM Fit rebrand, accessory recommendations, and deeper page analysis
Renamed the site to LocalLM Fit, refreshed the LM icon assets, added no-price accessory category links for relevant setups, and expanded model, hardware, VRAM, and Mac pages with quantization-specific fit tables, context ceilings, speed estimates, run commands, alternatives, and FAQs.
- Hardware matrix, curated combinations, Mac guides, and subscriptions
Expanded the hardware catalog with memory bandwidth and verification dates, added 50 curated model-and-machine pages, Apple silicon guides, Japanese/German routes, and a Cloudflare D1 notification form.
- Three-state calculator and VRAM guides
Added comfortable/tight/too-large thresholds, minimum comfortable hardware targets, five VRAM guides, FAQ structured data, and this methodology page.
- Expanded model library
Added source-backed GPT-OSS, Qwen, Gemma, and DeepSeek records with published quantization sizes, context, KV-cache estimates, runtime support, and license notes.
- Initial compatibility baseline
Established the first static model and hardware profiles, then validated the Cloudflare Pages build and deployment path.
Limitations
- Estimates are not runtime measurements or performance benchmarks.
- Operating-system reservation, display use, browser tabs, and other applications reduce usable memory.
- Longer contexts, concurrent requests, and larger batches can raise peak memory.
- Backend implementations can differ in workspace, KV-cache precision, and offload behavior.
- Price classes are broad editorial bands and are not live retailer quotes.
- Model licenses and source pages can change; verify the linked primary source before commercial use.
If a figure is wrong, send the page URL and a primary source to [email protected].