Instructions to use BaseIntelligence/top-prism-architecture with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use BaseIntelligence/top-prism-architecture with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="BaseIntelligence/top-prism-architecture", trust_remote_code=True)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("BaseIntelligence/top-prism-architecture", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use BaseIntelligence/top-prism-architecture with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "BaseIntelligence/top-prism-architecture"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BaseIntelligence/top-prism-architecture",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/BaseIntelligence/top-prism-architecture

SGLang

How to use BaseIntelligence/top-prism-architecture with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "BaseIntelligence/top-prism-architecture" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BaseIntelligence/top-prism-architecture",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "BaseIntelligence/top-prism-architecture" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "BaseIntelligence/top-prism-architecture",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use BaseIntelligence/top-prism-architecture with Docker Model Runner:
```
docker model run hf.co/BaseIntelligence/top-prism-architecture
```

top-prism-architecture

File size: 2,974 Bytes

---
library_name: transformers
pipeline_tag: text-generation
tags:
- prism
- custom-architecture
- trust-remote-code
- neural-architecture-search
license: apache-2.0
---

<div align="center">

![BASE Banner](https://github.com/BaseIntelligence/base/raw/main/assets/banner.jpg)

<h1 align="center">PRISM top architecture</h1>

<p align="center"><b>Global-best miner architecture on Base PRISM — benchmarks vs GPT-2 / GPT-2 Large</b></p>

</div>

---

## Benchmarks vs GPT-2 (Prism-protocol)

Prism-protocol **public** eval pack (1×RTX 5090). Accuracy: **↑ higher better**. BPB: **↓ lower better**. References: [GPT-2 (124M)](https://huggingface.co/openai-community/gpt2) · [GPT-2 Large (774M)](https://huggingface.co/gpt2-large) (eval-only; not miner trains).

| Metric | This model | GPT-2 | GPT-2 Large | vs GPT-2 | vs GPT-2 Large |
|---|---:|---:|---:|:---|:---|
| Val BPB (G1) | 3.7181 | 4.7595 | 4.1639 | ↓ -1.0414 ✓ better | ↓ -0.4458 ✓ better |
| HellaSwag | 0.360 | 0.355 | 0.395 | ↑ +0.005 ✓ better | ↓ -0.035 worse |
| ARC-Easy | 0.335 | 0.245 | 0.280 | ↑ +0.090 ✓ better | ↑ +0.055 ✓ better |
| ARC-Challenge | 0.295 | 0.240 | 0.280 | ↑ +0.055 ✓ better | ↑ +0.015 ✓ better |
| PIQA | 0.630 | 0.585 | 0.690 | ↑ +0.045 ✓ better | ↓ -0.060 worse |
| WinoGrande | 0.520 | 0.515 | 0.545 | ↑ +0.005 ✓ better | ↓ -0.025 worse |
| BoolQ | 0.630 | 0.575 | 0.640 | ↑ +0.055 ✓ better | ↓ -0.010 worse |
| LAMBADA | 0.955 | 0.970 | 0.985 | ↓ -0.015 worse | ↓ -0.030 worse |
| OpenBookQA | 0.310 | 0.320 | 0.335 | ↓ -0.010 worse | ↓ -0.025 worse |

### Compute notes

| | This model | GPT-2 | GPT-2 Large |
|---|---|---|---|
| Parameters | 107.0M | 124M | 774M |
| Size vs Large | 7.23× vs GPT-2 Large (774M) | 6.22× | 1× |
| Train tokens | — | _(eval-only)_ | _(eval-only)_ |
| Wall clock | 20539s | _(eval-only)_ | _(eval-only)_ |
| Sustained train throughput | — | n/a | n/a |
| GPU (harness) | `GPU 0: NVIDIA GeForce RTX 5090 (UUID: GPU-e31dbb89-6a01-a2fb-2684-3b6f0efb3f28)` | 1×RTX 5090 (eval) | 1×RTX 5090 (eval) |

Throughput ≈ `6 × N × D / wall` TFLOPS (dense transformer train FLOPs rule of thumb).

## Model card

| field | value |
|---|---|
| arch_id | `arch_f17d92b32a8c79f7` |
| bpb | `3.718067` |
| submission | `7b8658aedfdb158782567d09e26a9e819c578ff20c4db1d940e12de77cd9d6d2` |
| owner_hotkey | `462c4a7dfe30…` |
| hub repo | `BaseIntelligence/top-prism-architecture` |

## Load (trust_remote_code)

```python
from transformers import AutoModel, AutoConfig
cfg = AutoConfig.from_pretrained("BaseIntelligence/top-prism-architecture", trust_remote_code=True)
model = AutoModel.from_pretrained("BaseIntelligence/top-prism-architecture", trust_remote_code=True)
```

Weights: `checkpoint.pt` (Hub LFS when large). Load via `PrismCustomModel.from_pretrained` with `trust_remote_code=True`.

Companion GitHub publish (when configured) lives under `BaseIntelligence/prism` `top-model/`.