Following the widespread community adoption of the Qwen3.5 and Qwen3.6 series, we are pleased to introduce Qwen3.8, the most capable generation in the Qwen open-model family to date.
Runs On
What it takes to serve Qwen3.8-27B-NVFP4 (19.9B 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 | 39.7 GB | 47.7 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 8-bit | 19.9 GB | 23.8 GB | 1x MI300X (192 GB) Vultr |
$1.85 | 1x H100 $1.99 · 1x MI325X $2.00 |
| 4-bit | 9.9 GB | 11.9 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 Sep 18, 2026.
SAVRN's Notes on Qwen3.8-27B-NVFP4
Every precision row here lands on one MI300X with 192 GB at $1.85 an hour on-demand, so the hardware question is settled. Precision decides headroom: 47.7 GB of memory at 16-bit, 23.8 GB at 8-bit, 11.9 GB at 4-bit, and the rest of the 192 GB serves the 262,144 token context. Unsloth AI publishes it as an NVFP4 quantization of Qwen/Qwen3.8-27B, a dense vision-language model that reads images and video, built for coding, research and long agentic work.
Apache 2.0 terms let you run it commercially, modify it and redistribute it if you keep the notices and state your changes. Reconcile the numbers first: the name says 27B, the listing counts 19.9B parameters and the weight files total 23.4 GB; read the Qwen/Qwen3.8-27B base listing it was quantized from too. No Index host serves it by the token yet; the hourly rate is your only price.
Model Card
By Unsloth AI, published under apache-2.0, revision f0b7c9e722f5.
[!Note] This model works with vLLM. It also works with SGLang, but only with the v0.5.19, because the
lm_headis FP8 rather than 4-bit.
Read our How to Run Qwen3.8-27B Guide!
This NVFP4 quant uses Unsloth Dynamic V3.0 (preview) for SOTA quantization performance.
- Developer Role Support so Qwen3.8 can work in agentic tools like Codex and more!
- MTP for fast inference is available.
- Qwen3.8 can now be run and fine-tuned in Unsloth Desktop. Read our guide.
- Tool calling improvements: Makes parsing nested objects to make tool calling succeed more.
- See below for 4-bit Qwen3.8-27B run inside of Unsloth Desktop:
Qwen3.8-27B
Following the widespread community adoption of the Qwen3.5 and Qwen3.6 series, we are pleased to introduce Qwen3.8, the most capable generation in the Qwen open-model family to date.
Configuration
- Architecture
- Qwen3_5ForConditionalGeneration
- Context length (tokens)
- 262,144
- Layers
- 64
- Hidden size
- 5,120
- Feed-forward size
- 17,408
- Attention heads
- 24
- Key/value heads
- 4
- Head dimension
- 256
- Vocabulary size
- 248,320
- Model type
- qwen3_5
- Quantization
- compressed-tensors
Identity and Version
- Repository
- unsloth/Qwen3.8-27B-NVFP4
- Publisher
- Unsloth AI
- Task
- Not stated by the source
- Modality
- Other
- Library
- Not stated by the source
- Parameters
- 19.9B parameters
- Languages
- Not stated by the source
- Revision
- f0b7c9e722f5565102fff8481c99e4d86ae099c7
- First published
- 2026-08-13
- Last updated
- 2026-09-08
Files and Weights
13 files, 23.4 GB in total. The weights are 2 files totalling 23.4 GB in safetensors.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| model.safetensors | Weights | 22.6 GB | c473512c70ea |
| model_mtp.safetensors | Weights | 849.4 MB | 1d8268aa85ac |
| config.json | Configuration | 22.6 KB | — |
| generation_config.json | Configuration | 214 B | — |
| model.safetensors.index.json | Configuration | 164.4 KB | — |
| preprocessor_config.json | Configuration | 390 B | — |
| video_preprocessor_config.json | Configuration | 385 B | — |
| README.md | Documentation | 6.7 KB | — |
| chat_template.jinja | Other | 10.0 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
| tokenizer.json | Tokenizer | 20.0 MB | 06b9509352d2 |
| tokenizer_config.json | Tokenizer | 1.0 KB | — |
| vocab.json | Tokenizer | 6.7 MB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 23.4 GB
Released by Unsloth AI through its official repository on Hugging Face. Read the license.
Built From
- Derived from Qwen/Qwen3.8-27B
- Quantized from Qwen/Qwen3.8-27B
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 23.4 GB |
| 16-bit | 39.7 GB |
| 8-bit | 19.9 GB |
| 4-bit | 9.9 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About Qwen3.8-27B-NVFP4
How much GPU memory does Qwen3.8-27B-NVFP4 need?
About 47.7 GB at 16-bit and 11.9 GB at 4-bit: the weights (19.9B parameters) plus a working margin. A long context needs more.
What is the cheapest GPU to run Qwen3.8-27B-NVFP4 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 Qwen3.8-27B-NVFP4 commercially?
Yes. Qwen3.8-27B-NVFP4 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.
What is Qwen3.8-27B-NVFP4's context length?
262,144 tokens, from the maximum position embeddings in its published configuration.