Methodology
Every number on this site comes from the data files the site is built from and the rules on this page. We track 562 models, 38 labs and 53 hardware entries. Nothing is measured on real hardware yet, so memory and fit are estimates, not benchmarks.
Data sources and schedule
- Models come from the Hugging Face model API for each lab's organization: name, repository, task, parameter count, license, context length, KV cache shape and 30-day downloads. "Released" is the repository creation time on Hugging Face.
- API prices come from OpenRouter's public model list, one row per provider.
- Exchange rate: USD to INR from the ExchangeRate-API open access feed. The current rate is ₹94.29 per $1, fetched 14 Nov 2026.
- Hardware specs are entered by hand from vendor pages; each device page lists its sources.
- An automated ingest runs daily at 00:30 UTC and the site rebuilds after it.
Which models we include
For each lab we keep its 15 most downloaded open-weight repositories, plus up to 10 more released in the last 180 days with at least 5,000 downloads. We skip quantized copies and derivative repositories, and we never include safety-removed variants (repositories marked abliterated, uncensored, heretic or jailbreak).
Memory estimates
Weights use a fixed number of bytes per parameter, counting every parameter, including all experts in mixture-of-experts models:
| Format | Bytes per parameter |
|---|---|
| 4-bit (Q4) | 0.6 |
| 8-bit (Q8) | 1.06 |
| 16-bit (BF16) | 2 |
The KV cache is sized for an 8,192-token context at 16-bit: 2 × layers × KV heads × head size × 8192 × 2 bytes. When the model config does not give that shape, we use 8% of the weight memory instead.
The estimate is (weights + KV cache) × 1.1, a 10% allowance for runtime overhead, rounded to 0.1 GB. Longer contexts need more memory than shown. Image, video and speech models also need working memory that depends on resolution or clip length.
Usable memory and fit
- Apple Silicon Macs: 66% of unified memory counts as usable at 36 GB or less, and 75% above, because macOS and apps share the same memory.
- AI boxes such as DGX Spark: 85%, because the memory is shared with the operating system. A 128 GB box counts as 108.8 GB usable.
- Graphics cards: the listed memory, before runtime overhead.
A model fits when its estimate is at most 90% of usable memory. On the model pages, Tight means above 90% and up to 100%, and Too big means above 100%. Fit counts use 4-bit. Rumored devices never count as fitting, and models without a parameter count are left out of fit counts.
API prices
Providers charge separately for input and output tokens. To pick the cheapest offer we use a blended price: (3 × input + output) ÷ 4 per million tokens. Ranked by a 3:1 mix of input and output tokens. Ties are named on the model page. Input and output prices are always shown as well.
Prices are in US dollars per million tokens, with rupees beside them at the current rate of ₹94.29 per $1. Rupee amounts of ₹10 or more are rounded to whole rupees.
Runtime badges and commands
MLX and GGUF badges mean we found a repository with that format, not that we tested it. API means at least one hosted offer is recorded. Example commands on model pages are built from those repository ids and have not been run by us.
Lab logos
Logos are each lab's Hugging Face organization avatar. Where an avatar does not read at small sizes we show the lab's two-letter symbol instead. Logos identify labs; they do not imply endorsement.