---
license: apache-2.0
tags:
  - gguf
  - metacortex-ai
  - on-device
  - receipts
---

# metacortex-models

GGUF models used by [metacortex-ai](https://github.com/nicaibutou1993/metacortex-ai) for on-device AI with receipt-based attestation.

These files are hosted here to provide authoritative SHA-256 reference hashes for the [receipt model verification](https://github.com/nicaibutou1993/metacortex-ai/blob/main/docs/receipts-spec.md#model-verification-planned) feature.

## Models

| File | Parameters | Quantization | Size | Upstream Source |
|------|-----------|-------------|------|----------------|
| `Qwen3.5-2B-Q4_K_M.gguf` | 2B | Q4_K_M | 1.2 GB | [unsloth/Qwen3.5-2B-GGUF](https://huggingface.co/unsloth/Qwen3.5-2B-GGUF) |
| `Qwen3.5-2B-Q8_0.gguf` | 2B | Q8_0 | 2.0 GB | [unsloth/Qwen3.5-2B-GGUF](https://huggingface.co/unsloth/Qwen3.5-2B-GGUF) |
| `Qwen3.5-4B-Q4_K_M.gguf` | 4B | Q4_K_M | 2.6 GB | [unsloth/Qwen3.5-4B-GGUF](https://huggingface.co/unsloth/Qwen3.5-4B-GGUF) |
| `Qwen3.5-9B-Q4_K_M.gguf` | 9B | Q4_K_M | 5.3 GB | [unsloth/Qwen3.5-9B-GGUF](https://huggingface.co/unsloth/Qwen3.5-9B-GGUF) |
| `Qwen3.5-27B-Q4_K_M.gguf` | 27B | Q4_K_M | 16 GB | [unsloth/Qwen3.5-27B-GGUF](https://huggingface.co/unsloth/Qwen3.5-27B-GGUF) |
| `embeddinggemma-300m-qat-Q8_0.gguf` | 300M | Q8_0 | 313 MB | [ggml-org/embeddinggemma-300m-qat-q8_0-GGUF](https://huggingface.co/ggml-org/embeddinggemma-300m-qat-q8_0-GGUF) |

## SHA-256 Checksums

```
aaf42c8b7c3cab2bf3d69c355048d4a0ee9973d48f16c731c0520ee914699223  Qwen3.5-2B-Q4_K_M.gguf
1b04acba824817554f4ce23639bc8495ff70453b8fcb047900c731521021f2c1  Qwen3.5-2B-Q8_0.gguf
00fe7986ff5f6b463e62455821146049db6f9313603938a70800d1fb69ef11a4  Qwen3.5-4B-Q4_K_M.gguf
03b74727a860a56338e042c4420bb3f04b2fec5734175f4cb9fa853daf52b7e8  Qwen3.5-9B-Q4_K_M.gguf
84b5f7f112156d63836a01a69dc3f11a6ba63b10a23b8ca7a7efaf52d5a2d806  Qwen3.5-27B-Q4_K_M.gguf
6fa0c02a9c302be6f977521d399b4de3a46310a4f2621ee0063747881b673f67  embeddinggemma-300m-qat-Q8_0.gguf
```

## Purpose

Each metacortex-ai receipt includes a `gguf_sha256` field (SHA-256 of the local model file). Users can compare this against the hashes published here to verify the model file on disk is genuine and unmodified.

This does **not** prove that this specific model generated the response -- only that an unmodified copy of the model exists on the device. See the [receipts spec](https://github.com/nicaibutou1993/metacortex-ai/blob/main/docs/receipts-spec.md) for the full trust model.

## Usage with llama-server

```bash
# Chat model
llama-server --model Qwen3.5-9B-Q4_K_M.gguf --jinja --reasoning-format deepseek -ngl 99

# Embedding model
llama-server --model embeddinggemma-300m-qat-Q8_0.gguf --embedding --pooling mean -c 2048
```