Ruach Studio 2.0.1

The studio's own Writer on your machine, a take that plays before all of it has come on a slow line, Engine → Defaults, an https door for MIDI, and the fixes of the days after 2.0.0 (HERESY 1798–1811).

The Writer, on your own machine

• Qwopus3.5-9B-v3 (Jackrong's fine-tune of Qwen3.5-9B, Apache-2.0), quantized with AutoRound (W4A16, symmetric, group 32) and moved into GGUF bit for bit: the 4-bit weights as Q4_0, nothing quantized twice; text only.
• Two quants in goldhub/Ruach_Studio_Models_v2:
  • quality: 6.7 GB, about 7.5 GB of VRAM with a 12K context, for a 12 GB card;
  • compact: 5.5 GB, about 6.5 GB, for an 8 GB card.
• It runs through llama.cpp's llama-server, behind the studio's own /lab/llm/v1. There is no chat server to set up and no key, and nothing leaves the machine.
• It starts at the first question and leaves the card after ten minutes without one. It takes a card only when that card has room for all of it, else it runs on the processor.
• Engine → Writer chooses the quant, uses the model, or unloads it now.
• Where it is here and you never chose another chat server, the Writer asks it. Your own vLLM, LM Studio, Ollama or OpenRouter work as before.

Playing and choosing

• A take on a slow line (a studio far away) starts once its first minute is in and the rest will come in time. At home it waits for the whole take, as before.
• Engine → Defaults sets:
  • the ⏪ ⏩ step and the arrow keys' steps;
  • when a take starts playing;
  • what a click and a double-click do;
  • whether new takes play when done;
  • whether disliked takes show in the lists.
• Ratings and notes: a star right after a like; 🔥 and 💎 on every card after 👍; notes on workspaces, sections and LoRAs.
• Cards and workspaces: a double-click on a card plays or stops it; Move into… puts one workspace into another, as a subfolder or only its samples.

The rest

• An https door for MIDI on port 7443, because a browser gives Web MIDI only to a secure page. It uses the studio's own certificate, and Engine → Security decides who comes in.
• The server log:
  • its own window has every view;
  • the Trainer's progress shows in the log;
  • its buttons are icons.
• Models and caches: the studio's models never read from a cache, and the last of SDXL and Krea 2 is gone.
• Not on free cloud GPUs: the studio is too large for them.

Install / update

• New: follow INSTALL.md, or use Pinokio on Linux.
• From 2.0.0: use Engine → Updates in the page, or Pinokio's Update. The models you have stay.
• For the Writer's own model, use either:
  • bash launcher/extras.sh, which fetches the quant your card takes and builds llama-server (a few minutes);
  • Engine › Models and LoRAs › Studio writer, then launcher/extras.sh once for llama-server.

Known limits

• NVIDIA or the processor only: the Writer's model runs on an NVIDIA card or on the processor; on an Intel Arc it runs on the processor for now.
• A shared card: on a card the songs' engine also uses, the Writer may not find room while a song is rendering. It then runs on the processor, which is slower.
• From earlier releases: the DAW and the piano roll are still a beta, and the sound decode's ladder is still an experiment, as in 2.0.0.