Remory-Qwen3.8-27B is an open-weight model from Hanchen Xia. It has 1.9B parameters. At 16-bit it needs about 4.7 GB of GPU memory, which fits on 1x MI300X from $1.85 an hour, at the lowest prices in the SAVRN Index. It draws 49 downloads a month.
Remory supplements a context summary with learned residual memory, helping an agent retain information across context compaction. This release contains the Qwen3.8-27B-compatible residual compressor selected from our recent local evaluations.
Runs On
What it takes to serve Remory-Qwen3.8-27B (1.9B 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 | 3.9 GB | 4.7 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 1.9 GB | 2.3 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 1.0 GB | 1.2 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 19, 2026.
Remory-Qwen3.8-27B on every accelerator the SAVRN Index prices, at every precision
Model Card
Remory supplements a context summary with learned residual memory, helping an agent retain information across context compaction. This release contains the Qwen3.8-27B-compatible residual compressor selected from our recent local evaluations. An 88-second walkthrough of the method, real Qwen / GLM runs, and a paired comparison with context curves, repeated tool outputs, and errors. The checkpoint contains the memory module, which runs alongside the separately downloaded Qwen3.8-27B actor. It consumes source and summary hidden states and returns soft memory embeddings. Use the loader below; AutoModelForCausalLM.frompretrained does not load this module. model.safetensors and compressor.pt…
Excerpt from the card by Hanchen Xia.
Configuration
- Architecture
- SummaryResidualCompressor
- Stored precision
- float32
Identity and Version
- Repository
- mocoV3/Remory-Qwen3.8-27B
- Publisher
- Hanchen Xia
- Task
- Not stated by the source
- Modality
- Other
- Library
- pytorch
- Parameters
- 1.9B parameters
- Languages
- en
- Revision
- cb97f7b8d8ef196dd4902c3c13401ad7af55d975
- First published
- 2026-10-07
- Last updated
- 2026-10-09
Files and Weights
25 files, 15.5 GB in total. The weights are 2 files totalling 15.5 GB in pt, safetensors.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| compressor.pt | Weights | 7.8 GB | bb377765bbc0 |
| model.safetensors | Weights | 7.8 GB | 23d363c5aef7 |
| config.json | Configuration | 7.2 KB | — |
| context_compression/__init__.py | Configuration | 40 B | — |
| context_compression/grepseek_modeling.py | Configuration | 56.9 KB | — |
| context_compression/losses.py | Configuration | 17.4 KB | — |
| context_compression/qwen38_residual_modeling.py | Configuration | 10.8 KB | — |
| context_compression/qwen38_residual_serving_contract.py | Configuration | 13.5 KB | — |
| example.py | Configuration | 1.2 KB | — |
| manifest.json | Configuration | 707 B | — |
| media/remory-demo-notes.json | Configuration | 167.3 KB | — |
| remory.py | Configuration | 2.0 KB | — |
| README.md | Documentation | 3.8 KB | — |
| THIRD_PARTY_NOTICES.md | Documentation | 645 B | — |
| LICENSES/DFlash-MIT.txt | Other | 1.1 KB | — |
| LICENSES/Recursive-Memory-MIT.txt | Other | 1.1 KB | — |
| LICENSES/SpecForge-MIT.txt | Other | 1.1 KB | — |
| SHA256SUMS | Other | 2.2 KB | — |
| media/remory-demo-poster.png | Other | 125.9 KB | 9c49f082d867 |
| media/remory-demo.mp4 | Other | 7.1 MB | 7555139c034c |
| media/remory-demo.vtt | Other | 1.5 KB | — |
| paper/remory-arxiv-source.tar.gz | Other | 1.7 MB | c8af53db4e48 |
| paper/remory.pdf | Other | 1.7 MB | 13a9bdfd23e1 |
| requirements.txt | Other | 76 B | — |
| .gitattributes | Repository | 285 B | — |
License and Download
- License
- Not stated by the source
- Access
- Open weights, no gate
- Download size
- 15.5 GB
Released by Hanchen Xia through its official repository on Hugging Face.
Built From
- Adapter of Qwen/Qwen3.8-27B
- Derived from Qwen/Qwen3.8-27B
- Described by arXiv:2610.11287
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 15.5 GB |
| 16-bit | 3.9 GB |
| 8-bit | 1.9 GB |
| 4-bit | 1.0 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About Remory-Qwen3.8-27B
How much GPU memory does Remory-Qwen3.8-27B need?
About 4.7 GB at 16-bit and 1.2 GB at 4-bit: the weights (1.9B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run Remory-Qwen3.8-27B 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.