KINGY LOCAL AI / FREE TOOL
Your computer.
Your next AI model.
Find a model for your computer, verify a private document workflow, and decide whether an upgrade or cloud access makes sense.
START WITH YOUR HARDWARE
What can my computer run?
Calculations stay in your browser.
No account or hardware scan.
Loading the model catalogue…
How we calculate memory fit
We add pinned weight-file bytes, F16 KV cache for your context and session count, recurrent state where required, and runtime / workspace / allocator allowances. We subtract OS, app and display reserves from the relevant memory pool. “Comfortable” keeps at least 10% of the available pool free; “tight” means only the low estimate fits.
Dedicated GPU VRAM and system RAM are separate pools. Apple silicon and DGX Spark share one pool. This planner covers full GPU residency or CPU-only execution; partial GPU offload, swap, multi-GPU pooling and runtime-specific optimizations are not modeled. A memory estimate is not a compatibility test or a speed prediction. Hardware presets use catalogued capacity; use a custom value from your system for a closer estimate.
Ranking orders memory fit, then artifact size. It does not rank answer quality. Context limits and architecture parameters are pinned to the model revisions in the catalogue. Each result links its artifact and runtime sources.
ONE WORKSPACE / YOUR ASSUMPTIONS
Local versus cloud
Compare an existing machine or a new purchase with a cloud subscription or your total API spend. These editable figures are examples, not current provider prices or measured whole-system power.
30 days per month. Local monthly cost = hardware ÷ useful life in months + watts ÷ 1,000 × hours/day × 30 × electricity rate. Payback assumes all entered cloud spend can be replaced: purchase price ÷ (cloud spend − local electricity). Quality, usage limits, maintenance, idle power, taxes and resale value can change the decision. Our GPU-only power telemetry is not a wall-meter measurement.
Save a measured setup · Version 1.0.0
Qwen3-4B JSON extraction on a 24 GiB M4 Pro
Take the exact settings, prompts, answer keys and retained outputs with you. This setup uses Qwen3-4B Q4_K_M, llama.cpp build 10809 and a 4,096-token context on macOS 26.3. The measurements below were recorded September 7, 2026; this is a new export of that evidence, with no new model run.
- Invoice extraction: 3 of 3 accepted.
- Corrected owner and launch date: 3 of 3 accepted. Median request time: 0.687 seconds.
- Deadline boundary classification: 2 of 3 accepted.
- Price ceiling and consent filtering: 0 of 3 accepted.
All outcomes use the original strict JSON rules. Three attempts per task do not establish general reliability. Request time excludes loading, warmup and human repair. GPU placement and total unified-memory usage were not independently measured.
Download the setup, requests and evidence (ZIP) · Save the versioned setup (JSON)
The ZIP includes an offline verifier that re-scores the saved outputs. It does not download weights, start a model or call an API. The JSON is a setup record, not a community-result submission.
Inspect the matching hardware configuration · Read the original pilot and its limits
KINGY MEASURED / RETAINED RECEIPTS
Real speeds. Exact conditions.
These three recorded tests used Kingy’s physical RTX 4090 and Ryzen 9 7950X. Each number is the median of five 256-token synthetic decode samples with F16 KV cache. They do not predict speed on your selected setup, conversation latency or answer quality.
Verification and provenance
Receipt samples and timings are public below. The linked review attestations record a separate-path recomputation from retained raw logs. This release rechecked the published file hashes, timings and medians; it did not rerun model inference or independently remeasure the physical machine. The original raw log files are identified by hash in each receipt but are not bundled in the public release. Exact GGUF conversion lineage to the upstream base commit was not independently established.
llama-bench method · Kingy Local Lab methodology · Ollama context settings