Qwen3.8-27B Intel Arc Tuned

Qwen3.8-27B Intel Arc Tuned

Runs a 27B Qwen3.8 model in GGUF files retuned for Intel Arc GPUs, with pure decode near 22 tokens a second in the author's tests. For people running llama.cpp with SYCL on Arc cards.

Made by Greatjedi 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 GGUF weights of a 27B Qwen3.8 model, converted so they run fast on Intel Arc GPUs. The author says common 27B quantizations stall on Arc because the SYCL backend of llama.cpp lacks kernels for I-Quant tensors. The tuned files turn those tensors into Q4_K.

Who it suits: It suits people with an Intel Arc GPU, such as the Arc Pro B60 or the A770, who run language models locally with llama.cpp. The card lists English and Chinese.

Title: Who runs it
Industry | Who uses it | How they use it | Example
Software | Developers with Intel Arc cards | Run a 27B model locally with llama.cpp SYCL | A local coding assistant on an Arc GPU
Home labs | Local AI hobbyists | Get usable speed from an Arc card | Chat with a 27B model on an Arc Pro B60
Research | Quantization researchers | Study why I-Quants stall on SYCL | Compare the stock and tuned files

If you run an NVIDIA or Apple GPU, you can skip it. The card gives no figures for other hardware.

Where to use it:

• llama.cpp with the SYCL backend, on Intel Arc GPUs
• Arc Pro B60 with 24 GB, Arc A770 with 16 GB and Battlemage cards, according to the card
• The oneAPI SYCL runtime

Title: Jobs it does
Use case | Who | What they get back
Plain generation | Local users | Faster decode than the stock I-Quant files on Arc
Speculative decoding | Speed seekers | Over 41 tokens a second with the embedded prediction head

How to set it up:

• An Intel Arc GPU with enough memory for a file of about 14 to 15 GiB
• llama.cpp built with the SYCL backend
• Python, only to download the files with the Hugging Face tools

1. Install the Hugging Face Hub tools.
2. Download the repository files.
3. Load the tuned GGUF file in llama.cpp on the Arc GPU.

pip install -U huggingface_hub
hf download Greatjedi/Qwen3.8-27B-Intel-Arc-Tuned-GGUF

How to use it: Load a tuned file in llama.cpp with full GPU offload. The author measured speeds with 99 offloaded layers and flash attention on, on the Level Zero device. The newest file keeps an embedded prediction head, so speculative decoding needs no extra drafter model.

hf download Greatjedi/Qwen3.8-27B-Intel-Arc-Tuned-GGUF

Example: You load the tuned Q4_K file on an Arc Pro B60 and ask the model a question. In the author's tests, pure decode ran at about 22 tokens a second against 8.10 for a stock IQ3_XXS file. Your own speed will depend on your setup.

For newcomers and experts: If you are new, a quantized file is a shrunken copy of a model, and some shrinking styles run badly on some graphics cards. The first small thing to try is downloading one tuned file and loading it the way you load any GGUF.

If you know the field, Intel's XMX engines use DPAS, which needs dense integer types. The author says I-Quant codebooks need table lookups and have no SYCL kernels, so they fall back to the host. The tuned files convert them to Q4_K with 6-bit scales and keep RoPE, layernorms and attention projections. The trade-off is size: files are 13.72 to 15.29 GiB.

Find I-Quant tensors in the file
Convert them to Q4_K blocks
Keep sensitive tensors untouched
Load the file in llama.cpp SYCL

Good to know:

Title: Pure decode speed on an Intel Arc Pro B60
ISTA-DASLab IQ3_XXS, stock: 8.10 tok/s
Ridge, stock: 12.22 tok/s
Unsloth UD-Q4_K_S, stock: 12.97 tok/s
Ridge tuned Q4_K: 22.23 tok/s
DAS Lab tuned Q4_K: 22.36 tok/s
DAS Lab IQ3_S tuned MTP: 21.42 tok/s

• Benchmarks: all speeds come from the author's runs on one Arc Pro B60 with 24 GB, so treat them as a single data point.
• Speculative decoding: the card reports 41.28 tok/s and 41.88 tok/s on the embedded-MTP variants, with no external drafter memory.
• Prefill: about 245 tok/s on the Ridge tuned file.
• License: apache-2.0.
• Cost to run: no API key, account or paid service is named, but you need an Intel Arc GPU.
• Data it reaches: you download files from Hugging Face, and the card names no other access.
• Not for: other GPU brands, which the card does not cover.

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 26.9B
• Weights: about 13.4 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 7,481
• Likes: 31
• License: apache-2.0
• Languages: en, zh
• Based on: empero-ai/Qwen3.8-27B-Ridge-GGUF
• Tags: gguf, intel-arc, intel-arc-b60, battlemage, sycl, qwen, qwen3.8, quantized, speculative-decoding, hardware-accelerated, iq3-s, text-generation

Links

• Model: https://huggingface.co/Greatjedi/Qwen3.8-27B-Intel-Arc-Tuned-GGUF
• Files: https://huggingface.co/Greatjedi/Qwen3.8-27B-Intel-Arc-Tuned-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download Greatjedi/Qwen3.8-27B-Intel-Arc-Tuned-GGUF

Then follow the usage in its model card. Nothing in this kit runs by itself.