VelociQ8/PIXIE-Rune-v1.0-Q8A64-GGUF

Q8A64 model file for VelociQ8
Fast CPU inference for GGUF models on Arm and x86
Embeddings, dense and MoE LLMs
Runtime · Releases · Usage · All models

PIXIE-Rune-v1.0 · Q8A64

This repository holds PIXIE-Rune-v1.0 converted to Q8A64, the 8-bit weight format of VelociQ8, to run on the CPU. It turns text into vectors for semantic search, tuned for Korean and English and for the aerospace domain, with a prefix on queries only.

Base model	telepix/PIXIE-Rune-v1.0, by TelePIX
Type	Text embedding model, 1024-dimension vectors
Size	About 568M parameters (from the Hugging Face model metadata; the original card gives no count)
Context	6,144 tokens in the original model
Weights license	Apache License 2.0
Measurements	None of its own. Verified to load and run, see Measurements

This file loads only in VelociQ8 1.1 or later. Other GGUF tools, including standard llama.cpp, cannot load Q8A64 files. The automatic snippets that Hugging Face shows under Use this model (for llama.cpp, Ollama and similar tools) do not work with these files.

Get the runtime

VelociQ8 is free-to-use proprietary software built on llama.cpp. It ships the usual tools (llama-server, llama-cli, llama-embedding, llama-bench and llama-quantize) with the same options and the same HTTP API, including the OpenAI-compatible endpoints. Download the archive for your machine from the table (GitHub release) or from the runtime folder of this repository, which has the same files. Check it against SHA256SUMS from the same place.

Archive	For
velociq8-1.2-linux-arm64.tar.gz	Arm servers with dotprod: Ampere Altra, AWS Graviton2 and later, Neoverse N1 and later
velociq8-1.2-linux-arm64-multi.tar.gz	Any 64-bit Arm CPU; it picks the best kernels at startup
velociq8-1.2-linux-x64.tar.gz	x86-64 with AVX2 and FMA (Intel Haswell, AMD Zen and later)
velociq8-1.2-windows-x64.zip	Windows on the same x86-64 CPUs (tested on Windows 11)
velociq8-1.2-ollama-0.40.8-linux-arm64.tar.gz	Portable Ollama 0.40.8 for the CPU, Linux on Arm
velociq8-1.2-ollama-0.40.8-linux-x64.tar.gz	Portable Ollama 0.40.8 for the CPU, Linux on x86-64
velociq8-1.2-ollama-0.40.8-windows-x64.zip	Portable Ollama 0.40.8 for the CPU, Windows on x86-64

The Ollama archives are a portable Ollama for the CPU with VelociQ8 inside, and do not touch an installed Ollama; see Ollama. The Linux archives need glibc 2.28 or newer. There is also a Docker image for Linux on Arm and x86-64, ghcr.io/dhaern/velociq8 (see Docker). VelociQ8 runs on the CPU only, and its tools cannot download models: download the file from this repository first and pass it with -m.

Questions about this file go to the Community tab; problems with the runtime, to GitHub issues.

Files

File	Contents	Size
PIXIE-Rune-v1.0-Q8A64.gguf	Q8A64 weights	825.04 MiB
LICENSE	License text

SHA256 of the weights file, in the format sha256sum -c reads:

4b5daa4f8c43a5f3757e5a56a4e455a9c953a5973d2e32f46a6c193fd435c39b  PIXIE-Rune-v1.0-Q8A64.gguf

Quick start

Unpack the archive that matches your CPU (Linux Arm64 shown) and download the file:

tar xf velociq8-1.2-linux-arm64.tar.gz && cd velociq8-1.2-linux-arm64
REPO=VelociQ8/PIXIE-Rune-v1.0-Q8A64-GGUF
FILE=PIXIE-Rune-v1.0-Q8A64.gguf
curl -L -O https://huggingface.co/$REPO/resolve/main/$FILE

You can also download it from the Files tab or with hf download VelociQ8/PIXIE-Rune-v1.0-Q8A64-GGUF PIXIE-Rune-v1.0-Q8A64.gguf --local-dir .. On Windows use llama-server.exe from the zip; the options are the same.

Start an embeddings server. Use -t for the number of physical cores you want to give the model:

./llama-server -m PIXIE-Rune-v1.0-Q8A64.gguf --embedding -t 4 -b 2048 -ub 2048 --port 8080

Each input must fit in one batch: -b 2048 -ub 2048 accepts texts of up to 2,048 tokens (the default, 512, rejects longer ones). For longer texts raise both values, up to the context of the model; memory use grows with them.

Ask for the embeddings of a query and a document in one request (the prefixes are explained under Prompt):

curl -s http://127.0.0.1:8080/v1/embeddings \
  -H 'Content-Type: application/json' \
  -d '{"input": ["query: How do solar panels produce electricity?", "Solar panels use photovoltaic cells to convert sunlight into electrical energy."]}'

The reply has one vector of 1024 numbers per input, in the same order. Each vector has unit length, so the dot product of two vectors is their cosine similarity.

Prompt

PIXIE-Rune-v1.0 is used with a fixed prefix in front of the query text; documents are sent as plain text, with no prefix.

A query is sent like this (the prefix is everything before the sample question):

query: How do solar panels produce electricity?

A document is sent as it is:

Solar panels use photovoltaic cells to convert sunlight into electrical energy.

Measurements

This model has no measurements of its own. Results for other models are published per model family (embeddings, dense LLMs, mixture-of-experts LLMs) in benchmarks.md (speed and memory) and quality.md (accuracy against full precision).

Verified: loads with VelociQ8 1.2 and passes a basic functional test (2026-10-21).

What was checked, with the VelociQ8 1.2 runtime, on the file in this repository:

Check	Result
Vector size	1024
Reference	The original model, run with PyTorch on a CPU in 32-bit float
Texts	16: 8 queries and their 8 documents, in Spanish, English, French, German and Chinese
Cosine with the reference vectors, lowest / mean	0.99950 / 0.99959
Correct document ranked first	8 of 8 queries (reference: 8 of 8)

This is a functional check on a few short texts, not a quality benchmark. It shows that the converted file gives vectors very close to the original model's; it says nothing about retrieval quality on your own data.

About the base model

Source: the original card of telepix/PIXIE-Rune-v1.0, by TelePIX Co., Ltd. What follows is copied or condensed from it and is TelePIX's statement, not VelociQ8's; it leaves out the performance claims and the Python usage example. Read on 2026-10-21. If anything here differs from the original, the original is right. Benchmark tables are not copied: the nDCG@10 tables and the description of each benchmark are in the Quality Benchmarks section of the original card.

Description

PIXIE-Rune-v1.0 is an encoder-based embedding model trained on Korean and English information retrieval data, developed by TelePIX Co., Ltd. PIXIE stands for TelePIX Intelligent Embedding. The model is optimized for semantic retrieval in Korean and English, with a specialization in the aerospace domain: document understanding, technical QA and semantic search in aerospace and related high-precision fields. The card also reports results on open-domain Korean and English retrieval benchmarks.

Architecture and size

• Sentence Transformer model: an XLM-RoBERTa encoder, the vector of the first (CLS) token, then normalization.
• Maximum sequence length 6,144 tokens; 1024-dimension vectors; cosine similarity.
• The original repository's LICENSE says it is built with snowflake-arctic-embed-l-v2.0, the original model by Snowflake.

Languages

Multilingual, optimized for Korean and English.

Intended use

Semantic retrieval in Korean and English, in particular aerospace information retrieval: document understanding, technical QA and semantic search.

Limitations and risks

The original card does not say. It has no section on limitations or risks.

Prompt format

Queries are encoded with the query prompt of the model's configuration, query: ; documents are encoded without a prefix. The card's example says to use prompt_name="query" for queries.

License

The card says the model is licensed under the Apache License 2.0. The original repository has its own LICENSE file, which this repository copies unmodified: it starts with "Copyright (c) 2026 TelePIX Co., Ltd. All rights reserved.", says the model is built with snowflake-arctic-embed-l-v2.0 by Snowflake under the Apache License 2.0, and then gives the full Apache License 2.0 text.

Citation

@misc{TelePIX-PIXIE-Rune-v1.0,
  title={PIXIE-Rune-v1.0},
  author={TelePIX AI Research Team and Bongmin Kim},
  year={2026},
  url={https://huggingface.co/telepix/PIXIE-Rune-v1.0}
}

Licenses

Two licenses apply, and they do not mix:

• Runtime: VelociQ8 Software License Agreement. It covers the VelociQ8 software only, that is, the archives in the releases. The runtime is free to use for personal or commercial work.
• Weights: the license of the original model. The weights in this repository are a converted copy of the model described in About the base model. They stay under the original license, stated below. The VelociQ8 license agreement does not cover them and does not limit the rights that license gives you.

VelociQ8 is not affiliated with the developers of the original model, and this conversion is not endorsed by them.

License of this model

This repository contains a converted copy of telepix/PIXIE-Rune-v1.0. The weights were converted to Q8A64 and stored in a GGUF file. The model was not retrained or fine-tuned.

License: Apache License 2.0. The full text is in the LICENSE file of this repository.

_{Arm is a registered trademark of Arm Limited (or its subsidiaries) in the US and/or elsewhere. Other product and hardware names belong to their respective owners and are used only to say what VelociQ8 is compared against or measured on.}