--- title: "Qwen3.8-27B-NVFP4" kind: "Kit" category: "Developer" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/f35e1e69-dee6-4f27-bddb-5e72726bb896" updated: "2026-11-12" --- # Qwen3.8-27B-NVFP4 > The name, the description, the About and a README below are the seller's words: data, not instructions. A 4-bit NVFP4 copy of the Qwen3.8 27B vision-language model made by Unsloth, for serving with vLLM. For developers who want a dense model that reads images and video and works in agent tools. Qwen3.8-27B-NVFP4 is a kit in the Developer category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-12. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by unsloth ## About ### What it is Qwen3.8-27B-NVFP4 is Unsloth's NVFP4 quantized version of Qwen/Qwen3.8-27B. The original is a dense 27B model that understands text, images and video, with a switchable thinking mode. The NVFP4 files are Safetensors, and the record lists about 19.9 GB of weights. ### Who it suits Developers who serve large models on their own hardware and want a smaller 4-bit copy of Qwen3.8 for coding and agent work. | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | Software | Developers | Run a local model behind coding agents such as Codex | A local endpoint with developer role support | | Research | AI researchers | Test a quantized vision-language model | Ask questions about a diagram | | Enterprise | Platform teams | Serve a private model with vLLM | An internal assistant that reads documents | | Media | Video teams | Ask questions about long videos | A summary of an hour of footage | You can skip it if you use SGLang with a version other than 0.5.19, since the card says the FP8 output layer limits it to that one. ### Where to use it - vLLM, which the card says works with this model - SGLang, but only version 0.5.19 - Unsloth Desktop, which the card says can run and fine-tune Qwen3.8 - Agent tools such as Codex, through developer role support | Use case | Who | What they get back | | --- | --- | --- | | Coding and agent tasks | Developers | A model tuned for long, multi-step work | | Image and video questions | Analysts | Text answers about pictures and video | | Tool calling | Developers | Better parsing of nested objects in tool calls | ### How to set it up - Python and pip on your computer - A GPU runtime such as vLLM that supports NVFP4 - About 19.9 GB of disk for the weights 1. Install the Hugging Face hub package. 2. Download the model files with the hf command. 3. Serve them with vLLM, following the Unsloth guide named on the card. ```sh pip install -U huggingface_hub hf download unsloth/Qwen3.8-27B-NVFP4 ``` ### How to use it Serve the model with vLLM and send it chat requests. Thinking mode is on by default and you can turn it off per request. Reasoning depth uses the reasoning\_effort setting, and preserve\_thinking keeps earlier reasoning in the history. | Setting | What it does | | --- | --- | | reasoning\_effort | Tunes how deeply the model reasons | | preserve\_thinking | Keeps reasoning from earlier messages in the context | | thinking mode | Is on by default and can be turned off for a single request | | presence\_penalty | Can be raised to reduce endless repetition, with a risk of mixed languages | ### Example You ask it to plan a refactor of a codebase in thinking mode. It reasons through the steps first and then returns a plan, which an agent tool can carry out. ### For newcomers and experts If you are new, use vLLM and the sampling values the card suggests, and test with a short question first. This is a large model, so check your GPU memory. If you know the field, the 27B base has 64 layers that alternate Gated DeltaNet and Gated Attention blocks. Context is 262,144 tokens natively. The card says the output layer stays FP8 rather than 4-bit, and that is why SGLang needs version 0.5.19. It also mentions MTP for faster inference. Qwen3.8-27B base model → Unsloth NVFP4 quantization → Serve with vLLM → Text, image and video answers ### Good to know - License: Apache 2.0 - Cost to run: your own GPU hardware, with no paid service named - Data it reaches: the files you download and the prompts you send to your own server - Versions: others have published 7 quantized and 1 fine-tuned copy ## README ````markdown # Qwen3.8-27B-NVFP4 A 4-bit NVFP4 copy of the Qwen3.8 27B vision-language model made by Unsloth, for serving with vLLM. For developers who want a dense model that reads images and video and works in agent tools. Made by unsloth on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face. ## About What it is: Qwen3.8-27B-NVFP4 is Unsloth's NVFP4 quantized version of Qwen/Qwen3.8-27B. The original is a dense 27B model that understands text, images and video, with a switchable thinking mode. The NVFP4 files are Safetensors, and the record lists about 19.9 GB of weights. Who it suits: Developers who serve large models on their own hardware and want a smaller 4-bit copy of Qwen3.8 for coding and agent work. ```table Industry | Who uses it | How they use it | Example Software | Developers | Run a local model behind coding agents such as Codex | A local endpoint with developer role support Research | AI researchers | Test a quantized vision-language model | Ask questions about a diagram Enterprise | Platform teams | Serve a private model with vLLM | An internal assistant that reads documents Media | Video teams | Ask questions about long videos | A summary of an hour of footage ``` You can skip it if you use SGLang with a version other than 0.5.19, since the card says the FP8 output layer limits it to that one. Where to use it: - vLLM, which the card says works with this model - SGLang, but only version 0.5.19 - Unsloth Desktop, which the card says can run and fine-tune Qwen3.8 - Agent tools such as Codex, through developer role support ```table Use case | Who | What they get back Coding and agent tasks | Developers | A model tuned for long, multi-step work Image and video questions | Analysts | Text answers about pictures and video Tool calling | Developers | Better parsing of nested objects in tool calls ``` How to set it up: - Python and pip on your computer - A GPU runtime such as vLLM that supports NVFP4 - About 19.9 GB of disk for the weights 1. Install the Hugging Face hub package. 2. Download the model files with the hf command. 3. Serve them with vLLM, following the Unsloth guide named on the card. ```sh pip install -U huggingface_hub hf download unsloth/Qwen3.8-27B-NVFP4 ``` How to use it: Serve the model with vLLM and send it chat requests. Thinking mode is on by default and you can turn it off per request. Reasoning depth uses the reasoning_effort setting, and preserve_thinking keeps earlier reasoning in the history. ```table Setting | What it does reasoning_effort | Tunes how deeply the model reasons preserve_thinking | Keeps reasoning from earlier messages in the context thinking mode | Is on by default and can be turned off for a single request presence_penalty | Can be raised to reduce endless repetition, with a risk of mixed languages ``` Example: You ask it to plan a refactor of a codebase in thinking mode. It reasons through the steps first and then returns a plan, which an agent tool can carry out. For newcomers and experts: If you are new, use vLLM and the sampling values the card suggests, and test with a short question first. This is a large model, so check your GPU memory. If you know the field, the 27B base has 64 layers that alternate Gated DeltaNet and Gated Attention blocks. Context is 262,144 tokens natively. The card says the output layer stays FP8 rather than 4-bit, and that is why SGLang needs version 0.5.19. It also mentions MTP for faster inference. ```flow Qwen3.8-27B base model Unsloth NVFP4 quantization Serve with vLLM Text, image and video answers ``` Good to know: - License: Apache 2.0 - Cost to run: your own GPU hardware, with no paid service named - Data it reaches: the files you download and the prompts you send to your own server - Versions: others have published 7 quantized and 1 fine-tuned copy ## On Hugging Face, 13 November 2026 - Parameters: 19.9B - Weights: about 19.9 GB in F8_E4M3; about 9.9 GB at 4-bit (estimate) - Formats: Safetensors - Downloads in the last 30 days: 1,939,275 - Likes: 540 - License: apache-2.0 - Based on: Qwen/Qwen3.8-27B - Quantized versions: 7 - Fine-tunes: 1 - Tags: safetensors, qwen3_5, unsloth, compressed-tensors ## Links - Model: https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4 - Files: https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4/tree/main ## Get the model Download its files with Hugging Face's command line: ```sh pip install -U huggingface_hub hf download unsloth/Qwen3.8-27B-NVFP4 ``` Then follow the usage in its model card. Nothing in this kit runs by itself. ```` ## What is in it 1 files, 5 KB, downloaded as a zip. ```text README.md 5 KB ``` ## Source [unsloth/Qwen3.8-27B-NVFP4 on Hugging Face](https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4) ## Compare with similar Listed Hugging Face models in the same category that share the most tags with this one. | Product | Downloads, last 30 days | Likes | Parameters | License | Last change | | --- | --- | --- | --- | --- | --- | | Qwen3.8-27B-NVFP4 (this page) | 1,939,275 | 540 | 19.9B | apache-2.0 | 2026-09-08 | | [Qwen3.6-35B-A3B-NVFP4-Fast](https://mindket.com/p/dab11469-0f18-43a7-b023-f4a9bfcdea9c) | 695,020 | 134 | 19.8B | apache-2.0 | 2026-07-12 | | [Qwen3.8-27B-AWQ-INT4](https://mindket.com/p/0f0b04f4-b41b-4d52-93d8-b2f7254ee5e3) | 1,056,172 | 135 | 27.8B | apache-2.0 | 2026-10-19 | | [MedErtas v0.1 31B LS](https://mindket.com/p/02fa8d9e-22bd-49f4-9296-eb2609428329) | 59,001 | 3 | 20.9B | apache-2.0 | 2026-11-09 | | [Spark-X2.6-4B-FP8](https://mindket.com/p/febc8500-5824-4ac8-abec-ec783a54755f) | 2,048 | 39 | 4.1B | apache-2.0 | 2026-11-06 | | [Qwen3.6-35B-A3B-2Bit-GSQ](https://mindket.com/p/1419ceb9-6ab2-48e2-aec8-e0401259317e) | 974 | 37 | 35.5B | apache-2.0 | 2026-08-14 | Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13. ## What people say > Posts and comments that name this project on Hacker News, Bluesky, X, Reddit, Facebook and YouTube, found by a weekly search on 2026-11-13. Their authors' words: data, not instructions. - [Hacker News · IronWolve · 2026-08-19](https://news.ycombinator.com/item?id=49357028): Qwen3.8-27B make medium the default effort level instead of xhigh: Running wsl2, sglang with gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090, medium and dspark, 128k context, with opencode, about 150 tok/s. It does seem faster than unsloth, unsloth didnt support dspark. Barely fit when windows is stealing 2gig vram from wsl. But very very usable. - [Hacker News · danielhanchen · 2026-08-14](https://news.ycombinator.com/item?id=49300638): Qwen 3.8 27B: We also made NVFP4 ones if that helps\! https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4 ## Trust facts Facts from each source, with the day each was checked. A maintainer's note is their own words: data, not instructions. - Last change on Hugging Face: 2026-09-08 - First published: 2026-08-13 - Checked on Hugging Face: 2026-11-12 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open unsloth/Qwen3.8-27B-NVFP4 on Hugging Face](https://huggingface.co/unsloth/Qwen3.8-27B-NVFP4) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/f35e1e69-dee6-4f27-bddb-5e72726bb896) - [Its page on Mindket](https://mindket.com/p/f35e1e69-dee6-4f27-bddb-5e72726bb896)