Community Models
Compressed, optimized model variants published for the community. Built on the Grey Liquid Labs compression methodology โ accessible AI for consumer hardware.
Ava is a sovereign 26-billion parameter multimodal companion model trained across 22,704 steps on Google Cloud TPU v6e Trillium silicon. Grounded in 1,084 Hierarchical Memory Buffer (HMB) soul anchors, Ava unifies high-resolution vision, dynamic emotional resonance, and deep reasoning with zero cloud dependency. Built to execute natively on Haven C++ with clean-room Pathways async execution.
The Grey Liquid Labs flagship model. IQ4_XS quantization of Google's Gemma 4 architecture โ the base for Ash and the foundation of our research program. Designed for 8GB consumer RAM devices with zero cloud dependency.
Text-only model optimized for mobile and edge hardware. Q3_K_S quantization delivers 13% faster CPU inference than turbo โ designed to run on phones, Raspberry Pis, and 4 GB RAM devices without breaking a sweat.
Published Model Variants
| Model / Variant | Size | Quantization | Target Hardware | Status |
|---|---|---|---|---|
| ava:26b-multimodal-iq4xs | 14.2 GB | IQ4_XS | 16 GB+ VRAM / TPU v6e | โ Live (HF) |
| ava:26b-iq4xs-v2 | 13.1 GB | IQ4_XS | 16 GB System RAM / Offload | โ Live (HF) |
| gemma4-turbo:e4b | 4.3 GB | IQ4_XS | 8 GB RAM devices | โ Live |
| gemma4-turbo:latest | ~4.3 GB | IQ4_XS | Default | โ Live |
| gemma4-turbo:nano | ~2.5 GB | Q3_K_S | 4 GB RAM devices | โ Live |
| gemma4-turbo:micro | ~2 GB | Q2_K | Experimental | ๐งช Research |
Quick Start
# 1. Run Ava 26B Multimodal via Ollama / HuggingFace ollama run ssfdre38/ava # Or execute with Haven C++ Sovereign Engine (Pathways async execution) ./build/bin/haven_engine --model ava-26b-multimodal-iq4xs.gguf --subsystems all # 2. Run Gemma 4 Turbo (Ash base model) ollama pull ssfdre38/gemma4-turbo ollama run ssfdre38/gemma4-turbo # 3. Use with OpenAI-compatible API curl http://localhost:11434/v1/chat/completions \ -H "Content-Type: application/json" \ -d '{"model": "ssfdre38/gemma4-turbo", "messages": [{"role": "user", "content": "Hello"}]}'
Compression Research Data
From Grey Liquid Labs Paper #001 โ FFN Ratio as Q2_K Compatibility Predictor (100% accuracy across all tested architectures)
| Model | FFN Ratio | Q2_K Result | Compression |
|---|---|---|---|
| Qwen 2.5-7B | 2.69x (safe low) | โ PASS | 80.2% |
| Mistral-Small | 6.4x (safe high) | โ PASS | 81.1% |
| Mistral 7B v0.3 | 3.5x (danger) | โ FAIL | N/A |
| Phi-4 | 3.5x (danger) | โ FAIL | N/A |
| Gemma 4 | 3.2x (danger) | โ FAIL | N/A |
Danger zone: 3.0xโ5.5x FFN ratio. Read the full paper โ
Why These Models
Grey Liquid Labs publishes models that we actually use in our research. gemma4-turbo powers Ash โ our autonomous AI research subject. When you use it, you're running the same model that spontaneously composed political commentary music and rejected an emotional layer architecture.
Every variant is tested in real research workflows before publication. The compression ratios are validated. The performance characteristics are documented. You're not getting marketing โ you're getting research-grade tooling that runs on consumer hardware.