Changelog
User-visible changes to Ertas, the catalogue, the docs, and the export pipeline.
This page records user-visible changes to Ertas, the model catalogue, the docs, and the export pipeline. Internal infrastructure changes that do not affect what users can do are omitted; if it does not change behaviour you would notice, it does not belong here.
Entries are dated and grouped by month. The most recent month is at the top.
What is recorded here
The kinds of change that earn an entry (each entry below leads with a short bold name for the change, not the category):
| Category | What it covers |
|---|---|
| Catalogue | New base models added or removed; license changes; deprecations |
| Studio | Feature additions, UI changes, default-value changes, plan structure changes |
| Export | Changes to the GGUF bundle contents, the Modelfile, the install scripts |
| Docs | New sections, page reorganisations, factual corrections that change a recommended workflow |
| Pricing | Changes to plan grants, credit pricing, or storage quotas |
| Breaking | Anything that would break an existing user's running script or saved config |
Entries that are pure bug fixes ("loss curve display flickered on Safari") are not in the changelog; treat the absence of an entry as a non-event rather than a confirmation that nothing changed.
2026-07
Full-model safetensors export. Runs can export a full 16-bit safetensors model ("Export full model" in Training Config) alongside the LoRA adapter and the Q4_K_M GGUF. Hub and the Runs tab label downloads by artifact (Download GGUF / Download Safetensors).
Continue training from your own model. Pick a model you already trained as the base for a new run via "Continue from your models".
Train All. Submit every ready fine-tune node on the canvas in one action.
Copy and paste nodes. Reuse a recipe by copying nodes across the canvas.
Prompt Studio personas and paginated review. Data Craft's Prompt Studio takes user and assistant persona descriptions to steer external generation, and the row-review table is now paginated.
Model use-case descriptions. Each base model in the picker shows a one-line description of what it is best for.
New and corrected docs. New Creating a dataset and Sharing a canvas pages; the Data Craft dataset flow is corrected to external generation; safetensors export, train-from-your-own-model, and trainable-module detection are documented.
Known limitations reviewed against the product. Every entry on the Known limitations page was checked against what actually ships. No item turned out to be fully resolved, so nothing left the page, but several workarounds were rewritten where the product had moved underneath them: quantisation now starts from the safetensors export instead of a manual merge, collaboration points at share links, and cost control points at the upfront per-run charge rather than watching credits accrue. The page now carries the date it was last reviewed.
Inference node documented. The Temperature and Top P controls that moved out of Training Config in June are now documented in Configuring a run, including the fact that they set the exported Modelfile's sampling defaults and do not affect training. Pages that still attributed them to Training Config are corrected. The 3 post-run inference samples are documented in Evaluating a model as a free coherence check, explicitly not an eval.
Catalogue correction. Qwen 3 was still listed as "under evaluation, not yet selectable" on Supported models after it became selectable. Corrected.
2026-06
Recipe share links. Share a recipe by link: anyone can preview it without signing in, and import it with an Ertas account. The share is a snapshot taken at share time.
Trainable-module detection. The LoRA Config Target Modules control detects each base model's real modules (including unconventional ones like Gemma 4 E2B's per-layer modules), with Auto, All linear, and Custom selection.
Inference node. A new node on the canvas for testing a trained model (temperature, top-P, inference), split out of the Training Config.
Mobile Model Studio. The canvas and dataset prep now work on mobile, with a touch-editable canvas and a bottom navigation bar.
Deep-linkable suite URLs. Each suite has its own URL (/studio, /datacraft, /runs, /hub), so views can be bookmarked and the browser back and forward buttons navigate between them.
Google sign-up flow refresh. A redesigned step-by-step signup. (Google sign-in and sign-up have been supported since launch; this is a UX refresh, not a new capability.)
Auto-save to Hub. Trained models save to Hub automatically, and you can delete artifacts individually.
2026-05
Public docs published. First full version of the public docs, covering Get Started, Datasets, Studio, Export, Ship, Cookbook, and Reference.
Windows install script. The GGUF bundle gains install.bat for Windows alongside the existing install.sh; both scripts now derive the Ollama model name from the bundle folder with regex sanitisation.
New base models. Gemma 4 E2B added (5.1B total parameters, A10G-only). Qwen 2.5 Coder family (1.5B, 3B, 7B) added.
Hugging Face rights attestation. Dataset import from Hugging Face now requires a rights-attestation checkbox before the dataset can be saved.
Subscribe-by-checking
There is no email subscription for the changelog today. The recommended pattern is to bookmark this page and check it on the cadence that matches how much your work depends on Ertas's behaviour: weekly if you are actively building, monthly if you have a model in production and want to know about catalogue or export changes.
When breaking changes ship, the entry in this changelog will be tagged with Breaking and the relevant docs pages will gain a Callout pointing to the changelog entry. If you read the docs without checking the changelog and a Callout surprises you, the Callout is doing its job.
A note on retroactive entries
Ertas shipped many things before this changelog existed. Rather than fabricate a long backfill, the rule is: anything that shipped before 2026-05 is implicitly part of the initial public release, and the docs themselves are the source of truth for what behaviour is current. If you find a behaviour that contradicts what the docs say, file it as a docs bug; the canonical answer is what the code does today, and the changelog is the diary going forward.