Downloads · 30 days
9
1% of all-time downloads
sdragonehqo/dawkins-v1.1.5-awq
dawkins-v1.1.5-awq is a text generation model from sdragonehqo. Use it when you need the model to write or continue text. The card lists the license as llama3.3.
AWQ-quantized variant of Dawkins v1.1.5 for efficient GPU inference. Reduced from 140GB (BF16) to 40GB (4-bit) with minimal quality loss.
Downloads · 30 days
9
1% of all-time downloads
All-time downloads
651
Public
Parameters
70.6B
39.8 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors39.8 GB · 100%
How the weights are stored.
I3268.5B · 97%
From the Hugging Face model README
AWQ-quantized variant of Dawkins v1.1.5 for efficient GPU inference. Reduced from 140GB (BF16) to 40GB (4-bit) with minimal quality loss.
This is the production deployment model — the same Dawkins v1.1.5 fine-tune, compressed using Activation-aware Weight Quantization (AWQ) so it fits on a single 48GB+ GPU and runs fast on vLLM.
Currently deployed on RunPod Serverless vLLM:
https://api.runpod.ai/v2/v6oxqu5qt7ptpk/openai/v1 (OpenAI-compatible)import { DawkinsClient } from '@hqo/ai-client';
const client = new DawkinsClient({
apiKey: 'RUNPOD_API_KEY',
endpointId: 'v6oxqu5qt7ptpk',
});
const answer = await client.chat([
{ role: 'user', content: 'What is HqO?' },
]);
| Metric | BF16 (full) | AWQ (4-bit) |
|---|---|---|
| Size | ~140GB | ~39.8GB |
| Min GPU VRAM | 2x 80GB | 1x 48GB |
| Inference speed | Baseline | ~1.5-2x faster |
| Quality | Baseline | Negligible loss |
AWQ preserves the weights most important for activation patterns, so quality degradation is minimal compared to naive quantization.
Standard HuggingFace safetensors layout with AWQ quantization config in quantize_config.json.