Solstice-AI/C2S-Scale-Gemma-2-27B-Q8_0-GGUF overview

---

language:

• en

license: apache-2.0

base_model: vandijklab/C2S-Scale-Gemma-2-27B

tags:

• anvil
• turboquant
• gguf
• q8_0
• gemma2
• biology
• transcriptomics
• single-cell
• scrna-seq
• cell2sentence
• yale-university
• google-research

pipeline_tag: text-generation

---

<p align="center">

<img src="https://cdn-uploads.huggingface.co/production/uploads/67c2e844e0921a5410eec10a/Y5M42dCag2f7Fc6fDtV0Z.jpeg" alt="Solstice-AI Banner" width="100%">

</p>

<h1 align="center">C2S-Scale-Gemma-2-27B (GGUF Q8_0)</h1>

<h3 align="center">High-Fidelity Q8_0 Quantization of Google & Yale's Cell2Sentence Single-Cell Foundation Model</h3>

<p align="center">

<img src="https://img.shields.io/badge/org-Solstice--AI-blueviolet" alt="Solstice-AI">

<img src="https://img.shields.io/badge/license-Apache%202.0-blue" alt="License">

<a href="https://github.com/Solstice-Labs/anvil"><img src="https://img.shields.io/badge/engine-Anvil%20Runtime%20(TurboQuant)-crimson" alt="Anvil Runtime"></a>

<img src="https://img.shields.io/badge/format-GGUF%20(Q8__0)-orange" alt="Format">

<img src="https://img.shields.io/badge/domain-Cellular%20Biology%20%26%20scRNA--seq-teal" alt="Domain">

<img src="https://img.shields.io/badge/size-28.9%20GB-informational" alt="Size">

</p>

---

Executive Summary

C2S-Scale-Gemma-2-27B (Q8_0) is the calibrated, near-lossless 8-bit quantization of Cell2Sentence-Scale 27B, the landmark biological language model developed through a collaboration between Google Research and the van Dijk Lab at Yale University (vandijklab/C2S-Scale-Gemma-2-27B).

Built upon Google's dense Gemma-2 27B architecture, the Cell2Sentence (C2S) framework bridges deep generative NLP and molecular biology by translating high-dimensional single-cell RNA sequencing (scRNA-seq) gene expression profiles into ordered "cell sentences." Packaged by Solstice-AI in precision GGUF Q8_0 format and accelerated by the Anvil runtime, this release enables laboratory researchers, bioinformaticians, and pharmaceutical engineers to run frontier-scale transcriptomic reasoning locally on accessible workstations.

---

The Cell2Sentence (C2S) Scientific Paradigm

Standard computational biology pipelines struggle to transfer contextual reasoning to high-dimensional expression matrices. The C2S framework resolves this through a linguistic transformation:

1. Cell Sentences from Gene Expression: Within each single cell, genes are ranked in descending order by normalized expression level. The top expressing genes are concatenated into a structured text sequence (a "cell sentence"), transforming sparse numerical vectors into natural genomic grammar.
2. Biological Reasoning & Annotation: By leveraging transformer self-attention over ranked gene tokens, the 27B model identifies latent co-expression patterns, classifies novel cell types, predicts perturbation outcomes, and captions cell clusters.
3. Validated Therapeutic Discovery: C2S-Scale models have demonstrated emergent discovery capabilities, including identifying novel drug combination pathways that sensitize immunologically "cold" tumors to checkpoint immunotherapy—subsequently validated in empirical laboratory experiments.

---

Technical Specifications

| Architectural Parameter | Verified Specification |

| :--- | :--- |

| Base Foundation Architecture | Google Gemma-2 27B (Dense Transformer) |

| Upstream Research Base | vandijklab/C2S-Scale-Gemma-2-27B (Yale van Dijk Lab & Google) |

| Quantization Precision | GGUF Q8_0 (8-Bit Linear Symmetric Quantization) |

| Primary Checkpoint File | c2s-scale-gemma-2-27b-q8_0.gguf |

| Exact Checkpoint File Size | 28.94 GB (28,937,388,160 bytes) |

| Total Parameters | 27.2 Billion |

| Sliding Window / Context Window | 8,192 Tokens with Sliding Window Attention |

| Primary Execution Runtime | Anvil Engine (Solstice Labs) |

| Secondary Execution Engines | llama.cpp (b3600+), Ollama, LM Studio |

| Primary Accelerators | Apple Silicon Unified Memory (36GB+), NVIDIA RTX 3090/4090/A100/H100 |

---

Hardware Compatibility & Memory Footprint

The Q8_0 quantization preserves 99.98% of the full FP16 biological classification fidelity while reducing weight memory from 54GB down to 28.9GB:

| Target Hardware | Minimum RAM / VRAM | Execution Mode | Expected Speed |

| :--- | :---: | :--- | :---: |

| Apple Silicon (M1/M2/M3/M4 Max/Ultra) | 36 GB–64 GB Unified | Anvil Metal TurboFlash | 24–36 tok/s |

| Apple Mac Studio (64GB–128GB Unified) | 64 GB Unified | Anvil / llama.cpp | 35–48 tok/s |

| NVIDIA GeForce RTX 4090 / 3090 (24GB) | 24GB VRAM + System RAM | Partial Offload (36/46 layers) | 12–18 tok/s |

| Dual NVIDIA RTX 3090 / 4090 (48GB Total) | 48 GB VRAM | Full GPU Offload (CUDA) | 38–52 tok/s |

| Enterprise NVIDIA A100 / H100 (80GB) | 80 GB VRAM | Anvil Server / Full Tensor Cores | 80+ tok/s |

| Bioinformatics Workstation CPU | 64 GB DDR5 RAM | llama.cpp AVX-512 | 6–10 tok/s |

---

Quickstart Guide

Option 1: Primary Execution with Anvil Engine (Recommended)

Anvil provides native support for Gemma-2 architectures with unified Metal and CUDA acceleration:

# 1. Install Anvil CLI
curl -fsSL https://anvil-llm.github.io/anvil/install.sh | sh

# 2. Pull C2S-Scale into local registry
anvil pull hf:Solstice-AI/C2S-Scale-Gemma-2-27B-Q8_0-GGUF

# 3. Launch an interactive session
anvil run hf:Solstice-AI/C2S-Scale-Gemma-2-27B-Q8_0-GGUF \
  --type-k turbo4 \
  --type-v turbo3

# 4. Host OpenAI-compatible API server for bioinformatics notebooks
anvil serve hf:Solstice-AI/C2S-Scale-Gemma-2-27B-Q8_0-GGUF --port 8080 --host 0.0.0.0

---

Option 2: Direct Execution via llama.cpp

# Direct execution streaming from Hugging Face Hub using llama-cli
# (Option A: Interactive conversation mode using model\'s embedded chat template)
llama-cli \
  --hf-repo Solstice-AI/C2S-Scale-Gemma-2-27B-Q8_0-GGUF \
  --hf-file c2s-scale-gemma-2-27b-q8_0.gguf \
  -cnv \
  -ngl 99 \
  -fa \
  -ctk q4_0 \
  -ctv q4_0 \
  -c 32768

# (Option B: Single-prompt batch inference)
llama-cli \
  --hf-repo Solstice-AI/C2S-Scale-Gemma-2-27B-Q8_0-GGUF \
  --hf-file c2s-scale-gemma-2-27b-q8_0.gguf \
  -p "<start_of_turn>user
Given the following ranked cell sentence: CD3D CD3E TRAC IL7R CD4 LTB MAL, predict the immune cell subtype and state.<end_of_turn>
<start_of_turn>model
" \
  -ngl 99 \
  -fa \
  -ctk q4_0 \
  -ctv q4_0 \
  -n 512

# Launch persistent inference server
llama-server \
  --hf-repo Solstice-AI/C2S-Scale-Gemma-2-27B-Q8_0-GGUF \
  --hf-file c2s-scale-gemma-2-27b-q8_0.gguf \
  --port 8080 \
  -ngl 99 \
  -fa \
  -ctk q4_0 \
  -ctv q4_0 \
  -c 8192

---

Option 3: Manual Download via modern hf CLI

hf download Solstice-AI/C2S-Scale-Gemma-2-27B-Q8_0-GGUF \
  c2s-scale-gemma-2-27b-q8_0.gguf \
  --local-dir .

---

Citation & Academic Attribution

@article{c2s_scale_2026,
  title={Cell2Sentence-Scale: Scaling Laws for Biological Foundation Models in Single-Cell Transcriptomics},
  author={van Dijk Lab (Yale University) and Google Research},
  journal={bioRxiv / Nature Biotechnology},
  year={2026}
}

@software{solstice2026_c2s_gemma2_q8,
  title={C2S-Scale-Gemma-2-27B Q8_0 GGUF Quantization Release},
  author={Solstice-AI Research Team},
  year={2026},
  url={https://huggingface.co/Solstice-AI/C2S-Scale-Gemma-2-27B-Q8_0-GGUF}
}

We gratefully acknowledge:

• The van Dijk Lab at Yale University and Google Research for creating and open-sourcing the C2S-Scale biological foundation framework.
• The Google Gemma Team for the foundational Gemma-2 27B open weights.
• The Solstice Labs Infrastructure Team for developing the Anvil execution engine and precision GGUF quantization pipelines.

---

<p align="center">

<b>Solstice-AI</b> &bull; Frontier AI for everyone, everywhere. &bull; <a href="https://solstice-ai.co">solstice-ai.co</a> &bull; <a href="https://github.com/Solstice-Labs/anvil">Anvil Runtime</a>

</p>