Reviewed repository / hardware evidence

maple-preview: GPU & VRAM requirements

Maple Preview sits in an awkward hardware middle: the Hub exposes 20.21 billion safetensors parameters, which is far beyond small-model assumptions but below the familiar 32B tier. The repository is marked as an MIT-licensed preview and names MapleForCausalLM as its architecture. The practical decision is whether a 12 GB card is genuinely enough or merely close on paper.

See the precision plan → Inspect the source repository →

Decision summary

Where the memory boundary actually lands.

Its INT4 weight-plus-headroom floor is 11.30 GiB. A conservative 90% usable budget leaves only 10.8 GiB on a 12 GB card, so this page does not call 12 GB a fit. The first listed single-card tier that clears the floor is 16 GB. INT8 moves the first conservative tier to 32 GB, while BF16 needs an 80 GB tier before cache and runtime workspace.

Why this page exists. This page is distinct from generic 20B or 24B advice because it makes the 12 GB versus 16 GB boundary explicit using the repository's own parameter total and a declared usable-memory rule.

Weights first / precision by precision

The reproducible memory floor.

Every row uses 20.21B parameters × bits ÷ 8 ÷ 1024³, then adds 20% planning headroom. The single-card tier assumes 90% of labeled VRAM is available.

Mode	Bits	Raw weights	+ 20% floor	First single-card tier	24 GB GPU count*	Calculator
INT4	4 bits	9.41 GiB	11.3 GiB	16 GB	1	Edit plan →
INT8	8 bits	18.83 GiB	22.59 GiB	32 GB	2	Edit plan →
BF16	16 bits	37.65 GiB	45.18 GiB	80 GB	3	Edit plan →

*GPU count is capacity division against 90% usable VRAM. It does not prove the runtime can shard the model, avoid replication, or deliver acceptable interconnect performance.

INT4 card-fit matrix

Test the floor against common memory tiers.

Positive margin means the INT4 weight-plus-headroom floor fits inside the declared usable budget. It is not leftover context capacity: cache, runtime workspace, and system allocations are still unknown.

Card memory	90% usable	One-card result	Capacity-only count
8 GB	7.2 GiB	Short by 4.1 GiB	2
12 GB	10.8 GiB	Short by 0.5 GiB	2
16 GB	14.4 GiB	Clears by 3.1 GiB	1
24 GB	21.6 GiB	Clears by 10.3 GiB	1
32 GB	28.8 GiB	Clears by 17.5 GiB	1
48 GB	43.2 GiB	Clears by 31.9 GiB	1
80 GB	72 GiB	Clears by 60.7 GiB	1
96 GB	86.4 GiB	Clears by 75.1 GiB	1
141 GB	126.9 GiB	Clears by 115.6 GiB	1

Public source record

What the repository declares.

Metadata refreshed Aug 15, 2026. Open the source before relying on any field.

Open deepgrove/maple-preview →

Repository: deepgrove/maple-preview
Safetensors total: 20.21B parameters
Architecture: MapleForCausalLM
Model type: Not declared
Library: transformers
License metadata: mit
Access gated: No
Base model: Not declared in card metadata
Languages: en
Created: Aug 4, 2026
Last modified: Aug 4, 2026
Hub signal: 6.1K downloads · 360 likes · daily rank #6

Do not confuse capacity with deployment

Repository-specific caution.
Preview status is a reason to inspect the repository rather than infer maturity from the trending signal. The Hub metadata does not establish speed, runtime compatibility, prompt format, production stability, or a safe context length. A converted checkpoint can also differ from the full safetensors total used here, so replace the estimate with exact artifact bytes when available.

Formula boundary.
No KV cache, activations, graph capture, allocator fragmentation, offload, operating-system reserve, throughput, latency, or power is included.

Source boundary.
Hub trend, downloads, and likes help discover repositories. They are not a benchmark, endorsement, security review, or license opinion.

Before downloading weights

Three checks specific to this repository.

Treat the preview label as a prompt to verify the current model card, tokenizer, and runtime instructions before downloading any weights.
Do not force the 11.30 GiB INT4 floor onto a 12 GB card without measuring the runtime reserve that the capacity table cannot see.
Use the 16 GB tier as the first conservative single-card experiment, then add the intended context and concurrency to the calculator.

Passive backlink / model-card distribution

Let the repository carry its hardware evidence.

Add this passive SVG badge to a README, model card, documentation page, or benchmark report. It states the source-backed maple-preview INT4 floor and links to the assumptions—not to a download or quality claim.

Why this repository benefits. This badge is most useful beside a Maple Preview setup or benchmark because its link resolves the deceptive 12 GB near-fit: 11.30 GiB exceeds the conservative 10.8 GiB usable budget, making 16 GB the first reviewed INT4 tier.

Markdown for a README or model card

[![maple-preview INT4 floor 11.3 GiB - ResearchAudio](https://tools.researchaudio.io/badges/models/maple-preview-gpu-requirements.svg)](https://tools.researchaudio.io/models/maple-preview-gpu-requirements/?utm_source=model_badge&utm_medium=model_card&utm_campaign=ai_evidence_lab&utm_content=maple-preview-gpu-requirements#hardware-plan)

Badge visits use utm_source=model_badge, utm_medium=model_card, and this repository’s slug. Copying does not submit or store anything.

Nearest reviewed parameter totals

These pages use the same formula and source rules, making the capacity boundary comparable without pretending the models have equivalent quality or runtime behavior.

maple-preview FAQ

Three answers with the assumptions attached.

Use these as a first-pass hardware screen. Replace parameter arithmetic with exact artifact bytes and architecture-aware cache inputs before deployment.

Open the editable GPU calculator →

How much VRAM does maple-preview need?

maple-preview has 20.21 billion safetensors parameters in the public Hub metadata. The weight-plus-20%-headroom floors are 11.3 GiB at INT4, 22.59 GiB at INT8, and 45.18 GiB at BF16. These estimates exclude KV cache, activations, runtime workspace, fragmentation, offload, and speed.

Can maple-preview run on one GPU?

The first listed card tier whose 90%-usable budget clears the INT4 weight floor is 16 GB. That is an arithmetic starting point, not a runtime guarantee.

Does the Hugging Face signal prove maple-preview is good?

No. Downloads, likes, and trending position are discovery signals. They do not prove task quality, safety, speed, runtime compatibility, or licensing suitability.