dl-retrieval is an open-weight model from Stefan Becker, released under BSD 3-Clause License. It has 49,600 parameters and a 512-token context. At 16-bit it needs about 0 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.
Working implementation of CLIP for Retrieval using a small configuration. The repository focuses on transparent code and repeatable smoke tests; benchmark claims are deliberately omitted.
Runs On
What it takes to serve dl-retrieval (49,600 parameters): the memory its weights need at each precision, and the cheapest way to rent enough data-center GPUs to hold them.
| Precision | Weights | Memory needed | Cheapest setup | Per hour | Also fits |
|---|---|---|---|---|---|
| 16-bit | 0.0 GB | 0.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 0.0 GB | 0.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 0.0 GB | 0.0 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
Memory is the weights at that precision plus 20% for the runtime and a short context; a long context needs more. Prices are the lowest on-demand hourly rates in the SAVRN Index, read Oct 18, 2026.
dl-retrieval on every accelerator the SAVRN Index prices, at every precision
Model Card
By Stefan Becker, published under bsd-3-clause, revision a0eb53a854d5.
Working implementation of CLIP for Retrieval using a small configuration. The repository focuses on transparent code and repeatable smoke tests; benchmark claims are deliberately omitted. - The Python file contains the model and runnable example or training entry point. - config.json records the generated architecture settings. - trainingargs.json records the default experiment recipe. - model.safetensors is a valid initialization checkpoint for smoke tests; it is not presented as a trained benchmark checkpoint. - No benchmark score is claimed in this repository. The included configuration uses adafactor with a exponential schedule. These are starting values in the script, not evidence of a…
Read Stefan Becker's full model card
CLIP for Retrieval
Overview
Working implementation of CLIP for Retrieval using a small configuration. The repository focuses on transparent code and repeatable smoke tests; benchmark claims are deliberately omitted.
Repository status
- The Python file contains the model and runnable example or training entry point.
config.jsonrecords the generated architecture settings.training_args.jsonrecords the default experiment recipe.model.safetensorsis a valid initialization checkpoint for smoke tests; it is not presented as a trained benchmark checkpoint.- No benchmark score is claimed in this repository.
Architecture
| Item | Value |
|---|---|
| Architecture | CLIP |
| Scale | small |
| Attention | grouped query |
| Fusion | gated fusion |
| Activation | mish |
| Normalization | rmsnorm |
Default experiment recipe
The included configuration uses adafactor with a exponential schedule. These are starting values in the script, not evidence of a completed run. For a meaningful evaluation, train all baselines with the same data exposure, tuning budget, and random seeds.
Quick check
python predict.py --help
Inspect the script's __main__ block for its generated smoke-test example. Because this is a custom implementation, generic automatic loading APIs require an explicit adapter before use.
Evaluation guidance
A useful first evaluation would use Flickr30k, report the task metric across at least three seeds, and include a matched-capacity baseline. Keep training logs and environment versions with any published result.
Limitations
The initialization checkpoint has not been trained or audited for robustness, fairness, or domain transfer. The implementation should be treated as an experimental starting point. Results from a future trained checkpoint must be documented separately from the defaults shipped here.
Files
predict.py— primary artifactREADME.md— this documentationconfig.json— architecture configurationtraining_args.json— default experiment settingsmodel.safetensors— initialization checkpoint
License
Released under bsd-3-clause. Review the source-data terms separately when this repository is used with external datasets.
Configuration
- Architecture
- CustomResearchModel
- Context length (tokens)
- 512
- Layers
- 8
- Hidden size
- 384
- Feed-forward size
- 1,536
- Attention heads
- 2
- Model type
- clip
Identity and Version
- Repository
- stefanbeckerana/dl-retrieval
- Publisher
- Stefan Becker
- Task
- Not stated by the source
- Modality
- Other
- Library
- Not stated by the source
- Parameters
- 49,600 parameters
- Languages
- Not stated by the source
- Revision
- a0eb53a854d5560732f56e0d5725f425b07d0600
- First published
- 2026-10-10
- Last updated
- 2026-10-10
Files and Weights
6 files, 205.9 KB in total. The weights are 1 file totalling 198.9 KB in safetensors.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 198.9 KB | 1dd709633444 |
| config.json | Configuration | 427 B | — |
| predict.py | Configuration | 2.4 KB | — |
| training_args.json | Configuration | 198 B | — |
| README.md | Documentation | 2.4 KB | — |
| .gitattributes | Repository | 1.5 KB | — |
License and Download
- License
- bsd-3-clause
- Access
- Open weights, no gate
- Download size
- 198.9 KB
Released by Stefan Becker through its official repository on Hugging Face. Read the license.
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 198.9 KB |
| 16-bit | 0.0 GB |
| 8-bit | 0.0 GB |
| 4-bit | 0.0 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About dl-retrieval
How much GPU memory does dl-retrieval need?
About 0 GB at 16-bit and 0 GB at 4-bit: the weights (49,600 parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run dl-retrieval on?
At 16-bit, 1x MI300X from $1.85 an hour; at 4-bit, 1x MI300X from $1.85 an hour, at the lowest on-demand prices the SAVRN Index lists.
Can I use dl-retrieval commercially?
Yes. dl-retrieval is released under BSD 3-Clause License. The BSD 3-Clause License is permissive. It permits commercial use and redistribution with the copyright notice, and forbids using the authors' names to endorse derived products without permission.
What is dl-retrieval's context length?
512 tokens, from the maximum position embeddings in its published configuration.