This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| # Two DGX Spark recipes: Qwen 3.8 27B (SGLang + DSpark) + Ornith-1.5 35B-A3B (vLLM + MTP) | |
| Verified on 2x DGX Spark (128GB unified each). Both serve OpenAI-compatible endpoints. | |
| Recipes: r0b0tlab/qwen38-27b-nvfp4-sm121-sglang · MiaAI-Lab/Ornith-1.5-35B-A3B-DGX-Spark | |
| ## Spark 1 — Qwen 3.8 27B dense, NVFP4, SGLang with DSpark speculative decoding | |
| Weights: main model at `/model`, DSpark draft at `/draft` (in-container mounts). | |
| ```bash |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Every number in the H3 hardware landscape map, traced to its source. | |
| Consumer GPU benchmarks | |
| - kingy.ai H3 hardware review and guide (3060/4070/4090/5090 data, SGLang topology comparison) | |
| kingy.ai/ai/ai-guides/minimax-h3-local-installation-hardware-guide/ | |
| - ai-muninn.com RTX 5090 NVFP4 benchmark (coolthor pack, 175s measured) | |
| - xhinker Medium 5090 field report (~2.5 min T2V, ~7 min I2V, Blackwell INT8 vs FP8) | |
| - HuggingFace Comfy-Org/MiniMax-H3 discussion threads #3, #6, #11, #18, #26 | |
| huggingface.co/Comfy-Org/MiniMax-H3/discussions |