Pinned Loading

1. GLM-5.3-Flash-DFlash2-TP4-1M-Context GLM-5.3-Flash-DFlash2-TP4-1M-Context Public

   DFlash2 speculative decoding for GLM-5.3-Flash at TP4 across 4x DGX Spark (GB10/SM121), full 1,048,576-token context. Single-variable measurement against MTP-4: single stream 65.6 to 102 tok/s on c…

   Python

2. GLM-5.3-Flash-1M-Context-4x-DGX-Spark GLM-5.3-Flash-1M-Context-4x-DGX-Spark Public

   GLM-5.3-Flash at FULL 1,048,576-token context on 4x NVIDIA DGX Spark (GB10/SM121), TP=4. Breaks the 262,144 community ceiling via the vLLM null-block backport, 2.67x decode by dropping --enforce-ea…

   Python 2

3. hearth hearth Public

   Self-hosted observability for home AI compute clusters — vLLM / llama.cpp / SGLang / Ollama / LiteLLM, one pane of glass.

   JavaScript 6