Smallest weight footprints
Rank six baseline model profiles by calculated 16-bit weight storage, with explicit parameter scope, nominal counts, and serving-memory limitations.
- What is included
- The six baseline profiles with numeric parameter counts in our directory. This is not a ranking of all current models. New profiles with unresolved checkpoint scope are excluded.
- How this list is ordered
- Ascending 16-bit weight storage, in decimal GB. Equal values share a rank. Nominal parameter counts remain approximate.
Publisher sources are linked for every model. ≈ marks nominal parameter arithmetic. Weight storage excludes serving overhead. On smaller screens, scroll the table horizontally.
| Rank | Model | 16-bit weights | Context | Deployment note |
|---|---|---|---|---|
| 01 | ≈ 8 GB~4B parameters | 128K | A smaller weight footprint leaves more room for serving overhead. It does not establish a particular latency or quality level.Compare model | |
| 02 | ≈ 16 GB~8B parameters | 128K | Start with the workload you actually serve. A 128K model context does not mean a small GPU can serve 128K at your target concurrency.Compare model | |
| 03 | 16.4 GB8.2B parameters | 32K native | A useful small-model baseline. Start with a short context and measure the additional latency of thinking mode separately.Compare model | |
| 04 | ≈ 24 GB~12B parameters | 128K | The nominal 16-bit weights alone approach 24 GB. Leave room for runtime allocations, KV cache, and image processing.Compare model | |
| 05 | 65.6 GB32.8B parameters | 32K native | Weight precision changes the hardware shortlist dramatically. Quality and performance still need to be tested on the chosen checkpoint.Compare model | |
| 06 | 618 GB309B parameters | 256K | 15B active parameters does not mean 15B worth of weights. Size a fully resident deployment for the total parameter count.Compare model |
Read this as storage arithmetic
Each value is the profile's parameter count multiplied by two bytes. The ordering is based on those recorded counts, including rounded model sizes where marked. Actual checkpoint files can include different precision, additional modules, and quantization metadata. This calculation does not measure GPU allocations or confirm a supported 16-bit checkpoint.
The smallest entry is not automatically the best fit
Before selecting hardware, add the memory needed for your context length, concurrency, and runtime. Then validate quality and successful request completion. A 24 GB weight estimate can exceed a 24 GB GPU's usable serving budget. This list deliberately makes no 24 GB, 48 GB, or 80 GB fit claims.
Why newer models are missing
The current collection includes models whose total scope needs reconciliation across backbone, embeddings, encoders, and auxiliary heads. They stay outside this arithmetic ranking until that review is complete. The compact-model shortlist includes newer candidates without assigning them unverified memory figures.
Before you choose hardware
How much GPU memory do you actually need? explains the next planning step. For the distinction between publisher specifications, calculations, and measurements, read the BenchGrid methodology.