WhatModels — Public website¶

Development line: project:whatmodels · thread public-website
Last event: 2026-05-15 · 1 dated since 2026-05-15 · Researched: 2026-09-05 · confidence: medium

What it is¶

WhatModels is a model selector for users of Ollama, llama.cpp and other local runtimes to pick models by GPU or manual VRAM, estimating maximum context and generation speed.

• GPU catalog to choose hardware presets.
• Manual VRAM input to specify custom memory limits.
• Minimum context filters to screen out configurations that fall short.
• Quantization comparison to evaluate size against memory.
• Fit categories to group models into "runs well", "tight fit" and "doesn't fit".

Generation speed appears only for catalog GPUs and remains an estimate rather than a runtime benchmark.

The tool suits initial model and hardware selection, but production inference requires verification on local hardware.

Development line¶

• 2026-05-15 — WhatModels public website documented. On 2026-05-15, a message in the WhatModels thread linked to the project's public website. This establishes a dated public reference point for WhatModels, but the available evidence does not identify a launch, release, or feature change.

What changed¶

2026-05-15 — WhatModels operated as a web compatibility calculator for local LLMs. The published example configured two RTX 5060 Ti 16 GB cards, 16K context, a minimum of 10 tokens/s and 64 GB RAM.

After 2026-05-15 — We found no confirmed dated releases beyond the initial event. Current repository documentation describes a static browser calculation with no backend or API: VRAM combines model weights and KV-cache, and generation speed derives from memory bandwidth and weight size.

How to use this¶

As of 2026-05-15, practitioners can use the documented WhatModels public website as a project discovery point; its specific capabilities and release status still require separate verification.

1. Open the calculator and choose a GPU from the list or enter available VRAM manually. — https://whatmodelscanirun.com/
2. Set the minimum acceptable context; when selecting a GPU from the list, compare estimated tokens per second. — https://github.com/BenD10/whatmodels
3. Start with "Runs well" results; verify "Tight fit" options locally with the required context and quantization. — https://github.com/BenD10/whatmodels

Best practices¶

• Include KV-cache alongside weights: a model fitting by GGUF file size can run out of memory at the target context. — https://github.com/BenD10/whatmodels
• Treat speed estimates as preliminary: generation speed depends on memory bandwidth, quantization, context and runtime engine overhead. — https://github.com/BenD10/whatmodels
• Compare configurations on the same model and quantization at the same required context, then confirm the result in the target runtime. — https://habr.com/ru/articles/1035862/

Superseded by this¶

• 2026-05-16 — Choosing a local LLM solely by parameter count or file size is obsolete for long-context tasks: calculation must include KV-cache and quantization.

Still unknown¶

• No primary dated changelog for WhatModels exists for 2026-05-15 or later changes; the GitHub repository shows current code but provides no dated release milestones in the available material.
• The initial schema recorded the 2026-05-15 event without dedicated finding fields; outside this event, no subsequent dated development steps are confirmed.

Sources¶

Agent brief¶

• Subject: project:whatmodels, thread public-website, 1 dated events 2026-05-15 → 2026-05-15.
• Practical note: As of 2026-05-15, practitioners can use the documented WhatModels public website as a project discovery point; its specific capabilities and release status still require separate verification.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.