MaLoW-Qwen3-VL-4B-Last8s is an open-weight model from Zhuoming Liu, released under Apache License 2.0. It has 198M parameters. At 16-bit it needs about 0.5 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index.
Inference checkpoint for Memory as Weights: Internalizing Long-Term History for Streaming Videos. Paper: Memory as Weights: Internalizing Long-Term History for Streaming Videos Download the base model separately: Qwen/Qwen3-VL-4B-Instruct.
Runs On
What it takes to serve MaLoW-Qwen3-VL-4B-Last8s (198M 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 | 0.4 GB | 0.5 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 0.2 GB | 0.2 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 0.1 GB | 0.1 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 16, 2026.
MaLoW-Qwen3-VL-4B-Last8s on every accelerator the SAVRN Index prices, at every precision
Model Card
By Zhuoming Liu, published under apache-2.0, revision 6e2bae8cd8b4.
Inference checkpoint for Memory as Weights: Internalizing Long-Term History for Streaming Videos. Paper: Memory as Weights: Internalizing Long-Term History for Streaming Videos Download the base model separately: Qwen/Qwen3-VL-4B-Instruct. This repository contains inference modules, not a standalone base model. This variant uses the last 8 seconds as the current observation, represented by 8 frames. Follow docs/videoevaluation.md in the MaLoW code release for OVO-Bench and StreamingBench. The JSON metadata records the inference configuration; load it through the MaLoW runtime. Weights and inference metadata are distributed under Apache-2.0. See LICENSE and NOTICE; base models and benchmark…
Read Zhuoming Liu's full model card
MaLoW (Qwen3-VL-4B, Last 8 Seconds)
Inference checkpoint for Memory as Weights: Internalizing Long-Term History for Streaming Videos.
Paper: Memory as Weights: Internalizing Long-Term History for Streaming Videos
Repository: MaLoW.
Download the base model separately: Qwen/Qwen3-VL-4B-Instruct. This repository contains inference modules, not a standalone base model.
This variant uses the last 8 seconds as the current observation, represented by 8 frames.
Follow docs/video_evaluation.md in the MaLoW code release for OVO-Bench and StreamingBench. The JSON metadata records the inference configuration; load it through the MaLoW runtime.
Reported reference results (percent):
| OVO-Bench backward | Real-time | Forward | Average |
|---|---|---|---|
| 63.78 | 74.68 | 50.68 | 63.04 |
| StreamingBench average | Real-time | Omni | Proactive | SQA |
|---|---|---|---|---|
| 74.15 | 84.84 | 61.8 | 56.8 | 58.8 |
Weights and inference metadata are distributed under Apache-2.0. See LICENSE and NOTICE; base models and benchmark assets retain their respective upstream terms.
Identity and Version
- Repository
- zhuomingliu/MaLoW-Qwen3-VL-4B-Last8s
- Publisher
- Zhuoming Liu
- Task
- Not stated by the source
- Modality
- Other
- Library
- Not stated by the source
- Parameters
- 198M parameters
- Languages
- Not stated by the source
- Revision
- 6e2bae8cd8b442b31a8790dd285ab9754a68c328
- First published
- 2026-10-06
- Last updated
- 2026-10-16
Files and Weights
16 files, 802.0 MB in total. The weights are 2 files totalling 790.5 MB in safetensors.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| c2l_metadata.safetensors | Weights | 368 B | be598d139ac8 |
| model.safetensors | Weights | 790.5 MB | 23b4634c0726 |
| c2l_metadata.json | Configuration | 99.0 KB | — |
| processor_config.json | Configuration | 1.2 KB | — |
| release_manifest.json | Configuration | 2.4 KB | — |
| LICENSE | Documentation | 11.3 KB | — |
| NOTICE | Documentation | 1.3 KB | — |
| README.md | Documentation | 1.3 KB | — |
| LICENSES/SimplerEnv-MIT.txt | Other | 1.1 KB | — |
| LICENSES/StreamingBench-MIT.txt | Other | 1.1 KB | — |
| LICENSES/doc-to-lora-MIT.txt | Other | 1.1 KB | — |
| SHA256SUMS | Other | 1.2 KB | — |
| chat_template.jinja | Other | 5.3 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| tokenizer.json | Tokenizer | 11.4 MB | be75606093db |
| tokenizer_config.json | Tokenizer | 760 B | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 790.5 MB
Released by Zhuoming Liu through its official repository on Hugging Face. Read the license.
Built From
- Derived from Qwen/Qwen3-VL-4B-Instruct
- Described by arXiv:2610.22070
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 790.5 MB |
| 16-bit | 0.4 GB |
| 8-bit | 0.2 GB |
| 4-bit | 0.1 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About MaLoW-Qwen3-VL-4B-Last8s
How much GPU memory does MaLoW-Qwen3-VL-4B-Last8s need?
About 0.5 GB at 16-bit and 0.1 GB at 4-bit: the weights (198M parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run MaLoW-Qwen3-VL-4B-Last8s 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 MaLoW-Qwen3-VL-4B-Last8s commercially?
Yes. MaLoW-Qwen3-VL-4B-Last8s is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.