Reviewed repository / hardware evidence

fuse-1-Lite: GPU & VRAM requirements

fuse-1-Lite is a 5.72 billion-parameter repository whose card lists both LiquidAI/LFM2.5-2.6B and Qwen/Qwen3.6-35B-A3B as base models. That provenance makes it important not to infer memory from either parent name. The only number used for the weight floor is the 5.72 billion total exposed by this repository's own safetensors metadata.

See the precision plan → Inspect the source repository →

Decision summary

Where the memory boundary actually lands.

The calculated floors are 3.20 GiB at INT4, 6.39 GiB at INT8, and 12.78 GiB at BF16 after the shared 20% allowance. INT4 and INT8 both clear a conservative 8 GB card budget; BF16 first clears 16 GB. The smaller modes leave materially different room for cache and runtime overhead, so precision is still a deployment decision rather than a label.

Why this page exists. The standalone value is provenance-aware sizing: it prevents readers from applying the 35B parent name to a repository whose published safetensors total produces a very different memory floor.

Weights first / precision by precision

The reproducible memory floor.

Every row uses 5.72B 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	2.66 GiB	3.2 GiB	8 GB	1	Edit plan →
INT8	8 bits	5.33 GiB	6.39 GiB	8 GB	1	Edit plan →
BF16	16 bits	10.65 GiB	12.78 GiB	16 GB	1	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	Clears by 4 GiB	1
12 GB	10.8 GiB	Clears by 7.6 GiB	1
16 GB	14.4 GiB	Clears by 11.2 GiB	1
24 GB	21.6 GiB	Clears by 18.4 GiB	1
32 GB	28.8 GiB	Clears by 25.6 GiB	1
48 GB	43.2 GiB	Clears by 40 GiB	1
80 GB	72 GiB	Clears by 68.8 GiB	1
96 GB	86.4 GiB	Clears by 83.2 GiB	1
141 GB	126.9 GiB	Clears by 123.7 GiB	1

Public source record

What the repository declares.

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

Open Akahsizrr/fuse-1-Lite →

Repository: Akahsizrr/fuse-1-Lite
Safetensors total: 5.72B parameters
Architecture: Fuse3ForCausalLM
Model type: fuse3
Library: transformers
License metadata: apache-2.0
Access gated: No
Base model: LiquidAI/LFM2.5-2.6B, Qwen/Qwen3.6-35B-A3B
Languages: en
Created: Aug 5, 2026
Last modified: Aug 9, 2026
Hub signal: 2.3K downloads · 119 likes · daily rank #11

Do not confuse capacity with deployment

Repository-specific caution.
A merge or derivative can have runtime and prompt-format constraints that parameter arithmetic cannot reveal. The Apache-2.0 card metadata, Fuse3ForCausalLM architecture, and source lineage should be checked directly. This page does not infer compatibility, quality, or licensing obligations from either base model and does not recommend a converted artifact.

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.

Inspect how the repository combines its LFM2.5 and Qwen3.6 base models rather than assuming either parent's prompt format or memory plan applies.
Confirm that the deployment runtime implements Fuse3ForCausalLM; a comfortable weight floor cannot compensate for missing architecture support.
Choose between INT4 and INT8 based on exact artifacts and cache needs, since both clear 8 GB but leave very different working margins.

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 fuse-1-Lite INT4 floor and links to the assumptions—not to a download or quality claim.

Why this repository benefits. This badge prevents a provenance mistake on derivative-model pages: fuse-1-Lite reports a 5.72B total despite its LFM and Qwen parentage, so the linked 3.20 GiB floor comes from this repository rather than either parent name.

Markdown for a README or model card

[![fuse-1-Lite INT4 floor 3.2 GiB - ResearchAudio](https://tools.researchaudio.io/badges/models/fuse-1-lite-gpu-requirements.svg)](https://tools.researchaudio.io/models/fuse-1-lite-gpu-requirements/?utm_source=model_badge&utm_medium=model_card&utm_campaign=ai_evidence_lab&utm_content=fuse-1-lite-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.

fuse-1-Lite 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 fuse-1-Lite need?

fuse-1-Lite has 5.72 billion safetensors parameters in the public Hub metadata. The weight-plus-20%-headroom floors are 3.2 GiB at INT4, 6.39 GiB at INT8, and 12.78 GiB at BF16. These estimates exclude KV cache, activations, runtime workspace, fragmentation, offload, and speed.

Can fuse-1-Lite run on one GPU?

The first listed card tier whose 90%-usable budget clears the INT4 weight floor is 8 GB. That is an arithmetic starting point, not a runtime guarantee.

Does the Hugging Face signal prove fuse-1-Lite is good?

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