The ash ML workstation.

The workstation

• Pipeline graphs — training pipelines are authored on a blueprint-style canvas: datasets, transforms, models, losses, and schedules wired as nodes. Ash compiles the graph to native code through our toolchain; there is no Python environment, no dependency stack to babysit.
• The run screen — a live view of a training run: losses, diagnostics, and generated samples as they happen. Training is a first-class screen, not a log file.
• The experiment ledger — every run records its configuration, metrics, and artifacts. Any experiment can be reproduced, any two compared, any checkpoint recovered.
• Management & cache — datasets, model versions, and intermediate artifacts are managed and cached by the workstation, so long-running work resumes instead of restarting.

GPU-first training

Ash trains exclusively on the GPU — NVIDIA (Vulkan) and Apple (Metal), selected automatically — through the same GPU stack the engine renders with. The design goal is simple: your hardware, running flat out, with no cloud requirement and no per-run fees.

Trainer vs. runtime

Ash has two distinct surfaces, and only one of them is the paid product:

• The trainer (what you license) — the workstation above.
• The runtime — the execution primitives models need at inference time. The runtime is an installable substrate other software detects and uses; other BigBrain products ship the model artifacts they need and never require an ash purchase to run. If ash is present, they can use its runtime primitives as an enhancement.

With boltzmann

On a subscription that includes boltzmann, model development gains version control: experiments, model versions, and tickets flow through the same depots and streams as the rest of the project — automated model development and model version control in one loop.

Licensing & activation

• Ash activates with its own ASH-XXXX-XXXX-XXXX-XXXX key, separate from your godspeed key, on any supported OS, transferable between machines.

Status & roadmap

Ash is prealpha, in founders early access. The workstation shell, pipeline compilation, the run screen, and the experiment ledger exist and train our own production models today; packaged founder builds publish as the surfaces stabilize. Subscribers’ keys activate each build the day it publishes.