Blog
Engineering notes
Deep dives from the team building OpenAlchemy — inference internals, model optimization, and the open-source work that powers the platform.
- EngineeringOctober 18, 202612 min read
Inside our video stack: Wan 2.2 and MiniMax H3 on ComfyUI workers
How a /v1/videos request becomes a ComfyUI graph on a 32 GB GPU — checkpoints, frame math, distillation LoRAs, why the Wan 2.2 MoE rendered slow motion and the three-stage sampler that fixed it, and what an audio guide in H3 really does.
- ProductOctober 18, 20264 min read
More video on OpenAlchemy: Wan 2.2, lip-sync, and H3 with your own audio
Four new video models — Wan 2.2 speech-to-video, image-to-video and Animate, plus LatentSync — and MiniMax H3 clips that take an audio track. One keyframe, one API, a whole short film.
- ProductOctober 8, 20263 min read
Video generation with sound is live on OpenAlchemy
MiniMax H3 now runs on Grid compute: 4–15 second shots at 768P / 24 fps with native stereo audio, in text-to-video, image-to-video, and reference-to-video modes.
- EngineeringJune 9, 20267 min read
Whisper on RTX 50-series: our whisper.cpp fork and a production STT layer
Upstream whisper.cpp crashes on Blackwell GPUs. We fixed it — then wrapped Whisper in streaming, word-level timestamps, cancellation, and an OpenAI-compatible transcription API.
- EngineeringMay 30, 20269 min read
TurboQuant comes to llama.cpp: 2-bit and 3-bit KV cache compression
We forked llama.cpp to bring Google Research's TurboQuant to the KV cache — shrinking it ~5× at 3-bit while staying quality-neutral, with measured wins on Qwen.