— raw weight arithmetic plus 20% planning headroom.
Instrument 17 / source-to-model-card loop
Turn any public Hugging Face model into a transparent VRAM plan.
Paste the repository URL. This browser reads its public safetensors parameter total, calculates INT4, INT8, and BF16 weight floors, and gives the maintainer a passive Markdown badge that carries the source and assumptions with it.
Inspect a model →Public source lookup
Start with the repository, not the model name.
A model label can hide a derivative, quantization, merge, or different parameter total. The lookup accepts an owner/model ID or a full huggingface.co model URL.
Public metadata receipt
Model VRAM floor
Safetensors parameters—
Architecture—
Pipeline—
Library—
Access—
Last modified—
— raw weight arithmetic plus 20% planning headroom.
— raw weight arithmetic plus 20% planning headroom.
Boundary: these are parameter-derived weight floors. They do not include KV cache, activations, graph capture, allocator fragmentation, quantization metadata, CPU offload, runtime compatibility, interconnect, speed, quality, safety, or license interpretation.
Add architecture, context, GPU count, and usable VRAM →Powered-by distribution loop
Add the evidence path to the model card.
The image is a passive SVG. The destination reloads this exact public lookup with utm_source=hf_model_card_badge, so external model-card visits and downstream ResearchAudio signups can be measured separately.
What this tool proves
Repository identity.
The owner and exact model repository remain attached to the result.
Transparent arithmetic.
Parameters × bits ÷ eight, converted to GiB, then increased by 20%.
No quality verdict.
Download counts, model names, and memory fit do not prove output quality or runtime support.
Reusable distribution surfaces
Put the receipt where model teams already work.
The public Hugging Face Space creates a native discovery path. The GitHub Action turns model workflows into repeatable installation points. Both hand readers back to source-backed evidence with distinct attribution.
Native Hugging Face Space
Run the same public metadata receipt inside the Hugging Face ecosystem, with direct paths to the architecture calculator, tracked badge, and free technical briefing.
Static · no GPUOne versioned action
Add deepmehta11/researchaudio-hf-vram-action@v1 with the public Hugging Face model ID.
Evidence in every run
The job summary shows the public parameter total, raw arithmetic, configured headroom, source link, and explicit exclusions.
No model downloadChannel-level attribution
The Space uses huggingface_space, its model-card badge uses huggingface_space_badge, and the action uses github_action_badge, keeping adoption separate from search and paid traffic.
- uses: deepmehta11/researchaudio-hf-vram-action@v1
id: vram
with:
model: Qwen/Qwen2.5-7B-InstructOpen the Hugging Face Space → Install the free GitHub Action →
Source and method
The useful answer is smaller than a deployment promise.
The source is public Hub metadata. Hugging Face documents its model-info interface as the repository-information route, including optional expanded properties. This page requests only the public model ID, safetensors summary, architecture configuration, task pipeline, library, access state, and last-modified value. It does not download model weights. Review the official HfApi documentation for the current contract.
The total is a model-weight input. The safetensors total is converted into billions of parameters. For each displayed precision, the page multiplies the parameter count by four, eight, or sixteen bits, divides by eight to obtain bytes, converts bytes to binary GiB, and adds 20% planning headroom. The headroom makes the number less brittle, but it still does not transform a weight estimate into a measured runtime.
A repository can lack usable metadata. A GGUF-only project, unusual checkpoint layout, private repository, broken safetensors metadata, or missing public total cannot be sized safely from this API field. The tool fails closed and directs the reader to the manual calculator. It also refuses totals above two trillion parameters because a surprising metadata count needs checkpoint inspection before it becomes a hardware claim.
Mixture-of-experts models require resident-weight discipline. An active-parameter figure can describe how much of the network participates in a token's computation. It does not necessarily describe how many weights the serving system must store. This tool uses the repository's total safetensors parameters for the resident-weight floor and makes no claim about expert placement, offload, routing cost, or supported parallelism.
Exact artifacts can be larger or smaller. Real checkpoints can mix data types, include embeddings or duplicate tensors, add quantization scales, or publish a converted artifact whose byte size differs from ideal bit-width arithmetic. Hugging Face's own memory-estimator guidance describes model loading as one part of the memory decision. Inspect the exact files and runtime before purchasing hardware.
Context belongs in the next calculation. Autoregressive serving stores keys and values for prior tokens. The cache grows with layers, KV heads, head dimension, context length, concurrent sequences, and cache precision. Use the linked architecture-aware worksheet and the current Transformers KV-cache guidance after the source lookup.
Why the badge exists
Useful infrastructure can distribute the publication.
A badge is valuable only when it helps the model-card reader inspect a real decision. It should not imply endorsement, benchmark quality, compatibility, or a verified production recipe.
Paste the exact repository
The public API supplies the model identity and parameter total. Missing evidence stops the automatic result.
Source before sizingAdd one Markdown line
The passive badge can live in a GitHub README, Hugging Face model card, benchmark report, or deployment note.
No script in the badgeKeep attribution attached
Each destination includes the model slug, model-card source, shared campaign, and exact VRAM-plan anchor.
Visits before claimsHugging Face describes model cards as Markdown README files that support discovery, reproducibility, and sharing. The badge supplies a hardware-evidence link; it does not replace the model card's intended-use, limitation, training, dataset, or evaluation sections. See the official model-card guidance.
Hugging Face VRAM FAQ
Five boundaries keep the badge honest.
The fastest way to make a memory calculator misleading is to hide where the parameter count came from or what the number excludes.
How does the Hugging Face VRAM calculator get a model's parameter count?
It requests the public Hugging Face model API in your browser and reads the repository's safetensors total. Repositories without usable public safetensors metadata are not estimated automatically.
Does the result include KV cache and serving memory?
No. The three displayed results are model-weight arithmetic plus 20% planning headroom. KV cache, activations, runtime workspace, fragmentation, offload, and throughput remain separate deployment checks.
Can I add the generated badge to a Hugging Face model card?
Yes. Hugging Face model cards are Markdown README files. The generated passive badge links to a permalink that reloads the same public source lookup and attributes model-card visits.
Can I generate the VRAM receipt in GitHub Actions?
Yes. The free ResearchAudio Hugging Face VRAM Evidence Action reads public metadata, writes the calculation to the job summary, and emits a repository-attributed badge without modifying repository files.
Does ResearchAudio store the model ID I enter?
No. The lookup runs in the browser against the public Hugging Face API. The model ID appears in the page URL only when you load, copy, or share a model-specific permalink.
Evidence before infrastructure
Get the next model decision brief.
ResearchAudio turns release claims, model metadata, memory boundaries, and serving constraints into practical decisions for engineers and builders.
Prefer the hosted signup page?Join ResearchAudio free →