Image: docker pull avarok/atlas-alpha2
Hardware: NVIDIA DGX Spark GB10
| Model | HuggingFace ID | Type | MTP | tok/s |
|---|---|---|---|---|
| 35B MoE | Kbenkhaled/Qwen3.5-35B-A3B-NVFP4 | SSM+Attn+MoE | Yes | ~130 |
| 80B MoE | nvidia/Qwen3-Next-80B-A3B-Instruct-NVFP4 | SSM+Attn+MoE | Yes | ~100 |
| VL-30B | ig1/Qwen3-VL-30B-A3B-Instruct-NVFP4 | Attn+MoE (Vision) | No | ~100 |
| Nemotron-H 30B | nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 | Mamba-2+MoE+Attn | No | ~100 |
| 27B Dense | Kbenkhaled/Qwen3.5-27B-NVFP4 | SSM+Attn (Dense) | No | ~14 |
| 122B MoE | Sehyo/Qwen3.5-122B-A10B-NVFP4 | SSM+Attn+MoE | Yes | ~50 |
Download models: hf download <HuggingFace ID>
35B (recommended):
sudo docker run -d --name atlas --gpus all --ipc=host --network host \
-v ~/.cache/huggingface:/root/.cache/huggingface \
avarok/atlas-alpha2 serve Kbenkhaled/Qwen3.5-35B-A3B-NVFP4 \
--port 8888 --kv-cache-dtype nvfp4 --gpu-memory-utilization 0.88 \
--scheduling-policy slai --max-seq-len 8192 --max-batch-size 16 \
--speculative --mtp-quantization nvfp4
This project is licensed under the GNU Affero General Public License (AGPL). If you modify or deploy this image as part of a network-accessible service, you must make the complete corresponding source code available to its users under the same license. This applies whether you interact with the software directly or through a container. We chose AGPL specifically to ensure source availability in an era where binary-only distribution no longer provides meaningful obscurity.
For the full license text, see the LICENSE file included in this repository.
Content type
Image
Digest
sha256:880d25e54…
Size
1.8 GB
Last updated
7 months ago
docker pull avarok/atlas-alpha2