--- title: "Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF" kind: "Kit" category: "Research" price: "Free" seller: "Mindket" version: "1" url: "https://mindket.com/p/0792e045-08c2-4925-a170-f8d414b1a878" updated: "2026-11-13" --- # Qwen3.8-27B-GSQ-RCO-IQ3\_S-recovered-GGUF > The name, the description, the About and a README below are the seller's words: data, not instructions. Shrinks Qwen3.8-27B to a 3-bit IQ3\_S GGUF with retrained scales. A research file for people who study quantization, not a faithful stand-in for the original model. Qwen3.8-27B-GSQ-RCO-IQ3\_S-recovered-GGUF is a kit in the Research category, listed by Mindket. It is free. This is version 1, last updated on 2026-11-13. Download it on this page, a zip of its files. Nothing in it runs by itself. Made by Joakimpalm-Zen ## About ### What it is This repository holds an IQ3\_S GGUF of Qwen3.8-27B whose scale values were retrained by distillation. It starts from ISTA-DASLab's GSQ-RCO file, and only the scale fields in 399 tensors changed. The card calls it a research artifact, not a general release. ### Who it suits It suits researchers and tinkerers who study quantization and want to compare a scale-recovery experiment against the unmodified file. | Industry | Who uses it | How they use it | Example | | --- | --- | --- | --- | | Research | Quantization researchers | Compare the recovered file with the source file and the parent | Reproduce the card's KLD and perplexity readings | | AI hobbyists | Local model users | Run a small IQ3\_S file wherever the source file runs | Chat with a 27B model in a small footprint | | Tooling | Runtime developers | Test engine support for i-quant files | Check CUDA and Metal kernels on this file | | Education | Students | Learn how scales and codes differ in quantized blocks | Read the card's method and evidence | If you need a close copy of the original model, you can skip it, since the card says it is not one. ### Where to use it - Any engine that reads the source GGUF file - Runner version 0.5.3 or later on CUDA, and 0.5.6 or later on Metal, per the card - Windows, macOS or Linux, where you download the files with Python | Use case | Who | What they get back | | --- | --- | --- | | Quantization study | Researchers | A file with measured distance from the parent | | Local chat | Hobbyists | A small 27B model, with the limits the card states | | Speed test | Developers | Faster decoding with the added MTP draft head | ### How to set it up You need Python with pip and an engine that reads GGUF files. - Python and pip - About 13.4 GB of disk, as the 4-bit estimate suggests - A GGUF engine such as the Runner the card names 1. Install the Hugging Face client with pip. 2. Download the repository with the hf command. 3. Load the recovered GGUF file in your engine. If you want faster decoding, also load the MTP draft file. ```sh pip install -U huggingface_hub hf download Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF ``` ### How to use it Once loaded, you chat with it like any other GGUF model. The text shown here gives no run commands. The repository also holds an MTP draft file, which works as a speculative decoding companion. ### Example You load the file in a supported engine and ask a short question. You get a normal reply. The card warns that the model's next-token behavior differs from the parent, so judge it on your own prompts. ### For newcomers and experts This is an experiment, not a polished release. The first small thing to try is to read the limit notes on the card before you load it. Quantized blocks store integer codes plus fp16 scale fields. Here only the scale fields were retrained against the BF16 parent, while codes, format, byte length and offsets stay the same as the source file. The result sits closer to the parent than the source file, yet well outside the card's own fidelity bar. It also scores a lower perplexity than the parent, so its distribution is a different one. Qwen3.8-27B parent model → ISTA-DASLab IQ3\_S GSQ-RCO file → Retrain scale fields by distillation → Same codes and layout, new scales ### Good to know Limit first: the card says the file reads 0.4119 mean KLD against the float32 reference, about eight times its own bar. It is a learned derivative, not a faithful quantization. Mean KLD against the float32 reference, lower is closer | Label | Value | | --- | --- | | This file | 0.4119 | | ISTA-DASLab source file | 0.5616 | | Card fidelity bar, at most | 0.05 | Perplexity: on the card's reference corpus, this file reads lower than its parent. Perplexity on the reference corpus | Label | Value | | --- | --- | | This file | 21.54 | | ISTA-DASLab source file | 53.48 | | Parent model | 73.29 | - Earlier reading: on 500 short Wikitext-2 prefixes, the card reports a mean KLD of 0.0450. It says that holds for those positions only. - MTP head: the card says it adds about 1.9 times decode speed with llama.cpp on a GPU, and about 1.1 times with Runner 1.1.3 on CUDA, or 1.5 times on its int8 route. - License: Apache 2.0, as the License line says. - Cost to run: no API key, account or paid service is named. The author invites voluntary contributions. - Data it reaches: the prompts you give your engine, and Hugging Face for the download. - Not for: use as a replacement for the original model, since the card calls it a research file and not a general release. ## README ````markdown # Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF Shrinks Qwen3.8-27B to a 3-bit IQ3_S GGUF with retrained scales. A research file for people who study quantization, not a faithful stand-in for the original model. Made by Joakimpalm-Zen 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: This repository holds an IQ3_S GGUF of Qwen3.8-27B whose scale values were retrained by distillation. It starts from ISTA-DASLab's GSQ-RCO file, and only the scale fields in 399 tensors changed. The card calls it a research artifact, not a general release. Who it suits: It suits researchers and tinkerers who study quantization and want to compare a scale-recovery experiment against the unmodified file. ```table Industry | Who uses it | How they use it | Example Research | Quantization researchers | Compare the recovered file with the source file and the parent | Reproduce the card's KLD and perplexity readings AI hobbyists | Local model users | Run a small IQ3_S file wherever the source file runs | Chat with a 27B model in a small footprint Tooling | Runtime developers | Test engine support for i-quant files | Check CUDA and Metal kernels on this file Education | Students | Learn how scales and codes differ in quantized blocks | Read the card's method and evidence ``` If you need a close copy of the original model, you can skip it, since the card says it is not one. Where to use it: - Any engine that reads the source GGUF file - Runner version 0.5.3 or later on CUDA, and 0.5.6 or later on Metal, per the card - Windows, macOS or Linux, where you download the files with Python ```table Use case | Who | What they get back Quantization study | Researchers | A file with measured distance from the parent Local chat | Hobbyists | A small 27B model, with the limits the card states Speed test | Developers | Faster decoding with the added MTP draft head ``` How to set it up: You need Python with pip and an engine that reads GGUF files. - Python and pip - About 13.4 GB of disk, as the 4-bit estimate suggests - A GGUF engine such as the Runner the card names 1. Install the Hugging Face client with pip. 2. Download the repository with the hf command. 3. Load the recovered GGUF file in your engine. If you want faster decoding, also load the MTP draft file. ```sh pip install -U huggingface_hub hf download Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF ``` How to use it: Once loaded, you chat with it like any other GGUF model. The text shown here gives no run commands. The repository also holds an MTP draft file, which works as a speculative decoding companion. Example: You load the file in a supported engine and ask a short question. You get a normal reply. The card warns that the model's next-token behavior differs from the parent, so judge it on your own prompts. For newcomers and experts: This is an experiment, not a polished release. The first small thing to try is to read the limit notes on the card before you load it. Quantized blocks store integer codes plus fp16 scale fields. Here only the scale fields were retrained against the BF16 parent, while codes, format, byte length and offsets stay the same as the source file. The result sits closer to the parent than the source file, yet well outside the card's own fidelity bar. It also scores a lower perplexity than the parent, so its distribution is a different one. ```flow Qwen3.8-27B parent model ISTA-DASLab IQ3_S GSQ-RCO file Retrain scale fields by distillation Same codes and layout, new scales ``` Good to know: Limit first: the card says the file reads 0.4119 mean KLD against the float32 reference, about eight times its own bar. It is a learned derivative, not a faithful quantization. ```chart Title: Mean KLD against the float32 reference, lower is closer This file: 0.4119 ISTA-DASLab source file: 0.5616 Card fidelity bar, at most: 0.05 ``` Perplexity: on the card's reference corpus, this file reads lower than its parent. ```chart Title: Perplexity on the reference corpus This file: 21.54 ISTA-DASLab source file: 53.48 Parent model: 73.29 ``` - Earlier reading: on 500 short Wikitext-2 prefixes, the card reports a mean KLD of 0.0450. It says that holds for those positions only. - MTP head: the card says it adds about 1.9 times decode speed with llama.cpp on a GPU, and about 1.1 times with Runner 1.1.3 on CUDA, or 1.5 times on its int8 route. - License: Apache 2.0, as the License line says. - Cost to run: no API key, account or paid service is named. The author invites voluntary contributions. - Data it reaches: the prompts you give your engine, and Hugging Face for the download. - Not for: use as a replacement for the original model, since the card calls it a research file and not a general release. ## On Hugging Face, 13 November 2026 - Parameters: 26.9B - Weights: about 13.4 GB at 4-bit (estimate) - Formats: GGUF - Downloads in the last 30 days: 2,605 - Likes: 17 - License: apache-2.0 - Based on: ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF - Tags: gguf, qwen3.8, gated-deltanet, scale-recovery, distillation, xyntetik-runner, imatrix, conversational ## Links - Model: https://huggingface.co/Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF - Files: https://huggingface.co/Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF/tree/main ## Get the model Download its files with Hugging Face's command line: ```sh pip install -U huggingface_hub hf download Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF ``` Then follow the usage in its model card. Nothing in this kit runs by itself. ```` ## What is in it 1 files, 6 KB, downloaded as a zip. ```text README.md 6 KB ``` ## Source [Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF on Hugging Face](https://huggingface.co/Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF) ## 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-GSQ-RCO-IQ3\_S-recovered-GGUF (this page) | 2,605 | 17 | 26.9B | apache-2.0 | 2026-11-04 | | [Diffusion-Qwen3.8-27B](https://mindket.com/p/eda0723e-778a-4f43-807f-07f47bae3a29) | 2,015 | 33 | 26.9B | apache-2.0 | 2026-11-06 | | [Qwen3.5-4B-Qwenputer](https://mindket.com/p/cecaa8eb-29ad-4e59-8361-59c2f559a065) | 2,598 | 189 | 9.4B | apache-2.0 | 2026-11-05 | | [Qwenputer-npu](https://mindket.com/p/7f061dc6-4141-4afc-a196-0071e5354204) | 490 | 7 | 9.8B | apache-2.0 | 2026-10-30 | | [Sage-A1.5 GGUF (Q4\_K\_M)](https://mindket.com/p/7a186f44-d286-4d23-a115-416b5ad0f438) | 636,834 | 84 | 35.5B | apache-2.0 | 2026-10-22 | | [qwen-concept-27b](https://mindket.com/p/58083034-bfad-48bc-8a59-6417d3acb2ee) | 62,460 | 174 | 27.3B | apache-2.0 | 2026-11-08 | Numbers as Mindket last read them, from 2026-11-12 to 2026-11-13. ## 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: 2026-11-04 - First published: 2026-09-13 - Checked on Hugging Face: 2026-11-13 - Something wrong? Tell us: https://mindket.com/takedown#ask ## Get it - [Open Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF on Hugging Face](https://huggingface.co/Joakimpalm-Zen/Qwen3.8-27B-GSQ-RCO-IQ3_S-recovered-GGUF) - [Get it on your Mindket account](https://account.mindket.com/account/?product=mindket#/market/0792e045-08c2-4925-a170-f8d414b1a878) - [Its page on Mindket](https://mindket.com/p/0792e045-08c2-4925-a170-f8d414b1a878)