✨ Quark v0.1 released

Next-gen intelligence, scaled down.

Building highly efficient, logic-driven Small Language Models that run natively on edge hardware and consumer devices.

Available Models

Bilingual

Quark-135M 135M

A lightweight bilingual language model optimized for speed and localized logic. Click to expand variants.

Features GQA, SwiGLU, RMSNorm, and RoPE. Trained on 50B+ tokens of ultra-curated data.

Featured Model

Quark-270M 270M

Our scaled small model featuring 32 layers and 768 hidden dimensions for advanced reasoning capabilities.

Equipped with a dense 65K vocabulary. Specially designed for multi-turn instruct fine-tuning.

Legacy

Quark-50M 50M

Our compact 50M parameter model β€” engineered for extremely hyper-low resource systems.

Lightweight and highly volatile, ideal for basic sequence prediction and embedded units.

Explore Weights β†’
Instruct

Quark-72M ~72M

Instruction-tuned math & code model. Fine-tuned via SFT on a base checkpoint pre-trained on 5B tokens of math, code, and reasoning data.

14 layers, 512 hidden size, GQA (8Q/2KV), SwiGLU, RoPE. Pre-trained on 5B tokens, then SFT‑aligned. License: Apache‑2.0.

Explore Weights β†’
Safety Matrix

Quark-Mod Classifier

A high-throughput multi-label moderation engine covering 9 toxicity and cyber-safety categories.

Detects: toxic, severe_toxic, obscene, threat, insult, identity_hate, and advanced content exploits.

View Classifier β†’

Core Focus Areas

⚑

Hyper-Efficient

Mastering sub-1B parameters using Grouped-Query Attention (GQA) architectures.

🧠

Embedded CoT

Integrating step-by-step reasoning logic directly into the pre-training tokens.

🌍

Bilingual Focus

High-density blending of localized Italian, English, and technical STEM pipelines.

πŸ’»

True Open Source

All weights, configurations, and baseline streaming datasets are entirely open to the world.

Open Ecosystem Index

Repository Object Distribution Target
πŸ“š Quark-135M Base Foundational localized small baseline language matrix.
🌐 Quark-135M Bilingual Bilingual (IT + EN) checkpoint trained on balanced multi-source pools.
πŸš€ Quark-270M Instruct Multi-turn conversation alignment model with strict formatting safety.
βš™οΈ Quark-270M Base Base engine ready for specialized downstream task tokenization.
πŸ“¦ Quark-50M Legacy foundational checkpoint for exploratory sequence architectures.
πŸ›‘οΈ Quark-Mod Production safety guardrail classifier for modern pipeline filtering.
⚑️ Complete Collection (v0.1) Unified access hub to all current generation architectural releases.
πŸ’» GitHub Organization Training codebases, data streaming pipelines, and infrastructure layers.