# Models

Status: test-network preview. Prices are suggestions; each provider sets its own.

The curated catalog below makes the common case one click in the macOS app. The network is model-agnostic: an operator can also install other weights that fit the machine, and the `id` column is the value of the OpenAI `model` field.

| id | Name | Parameters | Quantization | Memory | Disk | License | Input credits / 1M | Output credits / 1M | Notes |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| `llama-3.2-3b-instruct` | Llama 3.2 3B Instruct | 3B | 4-bit | 4 GB | 1.8 GB | [Llama 3.2 Community License](https://www.llama.com/llama3_2/license/) | 0.02 | 0.04 |  |
| `qwen3-4b` | Qwen3 4B | 4B | 4-bit | 5 GB | 2.3 GB | [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) | 0.03 | 0.06 |  |
| `llama-3.1-8b-instruct` | Llama 3.1 8B Instruct | 8B | 4-bit | 7 GB | 4.5 GB | [Llama 3.1 Community License](https://www.llama.com/llama3_1/license/) | 0.05 | 0.08 |  |
| `qwen3-8b` | Qwen3 8B | 8B | 4-bit | 7 GB | 4.6 GB | [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) | 0.05 | 0.1 |  |
| `qwen3-14b` | Qwen3 14B | 14B | 4-bit | 11 GB | 8.3 GB | [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) | 0.08 | 0.2 |  |
| `gemma-3-27b-it` | Gemma 3 27B | 27B | 4-bit | 20 GB | 16 GB | [Gemma Terms of Use](https://ai.google.dev/gemma/terms) | 0.1 | 0.25 |  |
| `qwen3-30b-a3b` | Qwen3 30B A3B (MoE) | 30B total / 3B active | 4-bit | 21 GB | 17 GB | [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) | 0.08 | 0.25 |  |
| `qwen3-32b` | Qwen3 32B | 32B | 4-bit | 23 GB | 18 GB | [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) | 0.15 | 0.4 |  |
| `llama-3.3-70b-instruct` | Llama 3.3 70B Instruct | 70B | 4-bit | 46 GB | 40 GB | [Llama 3.3 Community License](https://www.llama.com/llama3_3/license/) | 0.3 | 0.6 |  |
| `qwen3-235b-a22b` | Qwen3 235B A22B (MoE) | 235B total / 22B active | 4-bit | 145 GB | 132 GB | [Apache-2.0](https://www.apache.org/licenses/LICENSE-2.0) | 0.4 | 1.2 |  |
| `muse-spark-open` | Muse Spark Open | 816B total / 42B active | NVFP4 | 560 GB | 493 GB | [Muse Model License](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | 1.25 | 4.25 | cluster only |

## Columns

- **Memory** is the unified memory the model needs once loaded, including working headroom. A Mac can give a model about two thirds of its memory under 64 GB and about three quarters above that.
- **Disk** is the approximate download size.
- **Price** is the suggested price in Keeta Compute Credits (one credit is one US dollar) per million input and output tokens. On the wire, prices are integer base units: credits have 6 decimals, so 0.05 credits is `50000`. [Pricing math](https://keeta.ai/compute/protocol.md).
- **Cluster only** marks a model no single machine can hold. Clusters are not scheduled in this version.
- Live prices and provider counts come from `GET /v1/models` on the [router](https://keeta.ai/compute/router.md) or the app's local endpoint.

## Licenses

License terms travel with the model. An operator who offers a model to the network is serving it commercially and takes on that license's obligations, such as attribution ("Built with Llama"), acceptable-use policies, and size or revenue thresholds that need a separate license. Those obligations are the operator's responsibility, not the network's. The app shows each license before you offer the model; read it in full at the linked page.

Every receipt names the exact weights the provider ran by their sha256 hash, so a buyer can tell which build of a model served the request.
