GLM-5.3-Flash BIZ serves a 1 million-token context on DGX Spark systems via an OpenAI-compatible API. Decode speed reaches 53.78 tok/s at TP=2 and 68.52 tok/s at TP=3. The new Jammed XSFP4 checkpoint reduces memory use by 6.58 GiB per host at TP=2.
Key points
Decode speed: 53.78 tok/s for TP=2, 68.52 tok/s for TP=3
Jammed XSFP4 saves 6.58 GiB memory per host at TP=2
Sources
- Original post
- GLM-5.3-Flash-DGX-Sparks-BIZ/v3 at main - Linked page
Pixelorama – open source pixel art editor with web version
Pixelorama is an open-source pixel art tool that runs on Windows, Linux, macOS and the web. It adds a 3D layer, a command-line interface for bulk exports, and supports PNG, animated PNG, sprite sheets, GIF and video export.
Key points
Runs on Windows, Linux, macOS and the web.
Animation timeline includes onion skinning, audio sync, frame tags and drawing while playing.
Sources
- Original post
- Orama-Interactive/Pixelorama - Linked page
Axolotl adds Mistral Large 4, QLoRA export to GGUF
Axolotl adds Mistral Large 4 support with merge-aware NVFP4 LoRA and enables GGUF export with optional quantization. The framework also offers CPU-staged FSDP2 QLoRA loading for models larger than a single GPU’s memory.
Key points
Mistral Large 4 supported with merge-aware NVFP4 LoRA and expert parallel training
axolotl export writes GGUF with optional quantization for llama.cpp and Ollama
Sources
OfficeCLI – single binary adds AI control of Office files
OfficeCLI is a single binary that lets AI agents create, read and edit Word, Excel and PowerPoint files without an Office installation. It renders .docx, .xlsx and .pptx to HTML or PNG for visual feedback. The tool reduces a task that required 50 lines of Python and three libraries to one command.
Key points
Single binary with no Office install needed.
Renders each document to HTML or PNG for the agent to view and fix.
Replaces a 50-line Python script and three libraries with one command.
Sources
- Original post
- OfficeCLI - Linked page
