Supported models

This page is the authoritative list of what Ertas can fine-tune: 16 base models, and the model picker in Studio shows exactly these. A model that is not on this page is not selectable, and a run submitted against one is rejected. The models index is the marketing view of the wider open-model landscape, with license excerpts and taxonomy, and is not a list of what you can train here.

If the model you want is not below, that is not the end of the road: the bring-your-own path takes any compatible text model from a Hugging Face URL.

The catalogue

GGUF sizes are at Q4_K_M (the only quantisation Ertas exports today, see Quantization). The GGUF is not the only artifact a run can produce: the LoRA adapter and, on paid plans, a full 16-bit safetensors model are separate downloads, and both are much larger than the GGUF column suggests. See File sizes and formats. "Free plan?" answers whether a Free-plan account can train this model; A10G models require Builder or higher. Sizes marked with ~ are estimates from the parameter count rather than measurements of a published GGUF.

Llama family

Mistral family

Mixtral is not in the catalogue. Mixture-of-experts models are not blocked on architecture, but training holds every expert in memory rather than just the active ones, so Mixtral 8x7B's 47B total parameters put it far outside what our GPU tiers handle. See Known limitations.

Phi family

Gemma family

Two Gemma entries need A10G for reasons that are not obvious from their parameter counts. Gemma 4 E2B is 2.3B effective but 5.1B total once its Per-Layer Embedding tables are counted. Gemma 3 4B is only 4.33B, under the usual 5B threshold, but Gemma 3 cannot train in bf16 on a T4 (Turing has no bf16 support and fp16 overflows for this family), so it falls back to fp32 at roughly double the memory and exceeds the card. Both deploy fine at 3B-class sizes once exported; the constraint is training, not inference.

Qwen family

LFM2.5 family (Liquid AI)

Purpose-built edge models, and the smallest bases in the catalogue. If you are targeting a phone, a browser, or anything memory-constrained, start here rather than shrinking a larger model down.

All three carry a 32,768-token context window, which is long for models this small. (Liquid AI publishes 32,768 for LFM2.5 230M, 350M and 1.2B alike.)

Under active evaluation (not in the catalogue)

Models in this group are being tested for catalogue inclusion. They are not selectable in the model picker, and a run submitted against a model outside the catalogue is rejected. You can still train them today through the Hugging Face URL import path, which is a separate path and does not go through the catalogue, at the cost of giving up automatic credit refunds on training failures.

Qwen 3.5 4B has graduated out of this group and is in the catalogue above. Other Qwen 3 variants are not.

Picking from the catalogue

If you are deciding between models, the right path is Picking a base model. Three quick heuristics that cover most cases:

• For mobile, default to a 3B-class model at Q4_K_M (Llama 3.2 3B, Qwen 2.5 3B, Phi-4 Mini, Gemma 3 4B). Below 1B is rough for instruction following; above 3B starts to crowd phone RAM.
• For desktop, an 8B-class model (Llama 3.1 8B, Mistral 7B, Qwen 2.5 7B) is a noticeable quality bump over 3B and still fits in 5 GB on disk.
• For web and tight memory budgets, prefer the LFM2.5 family or 1B-class models (LFM2.5 350M or 1.2B, Llama 3.2 1B, Gemma 3 1B, Qwen 2.5 1.5B) to keep first-load downloads small and stay inside browser memory ceilings.

Hugging Face models outside the catalogue

You can fine-tune any compatible text model from a Hugging Face URL. Ertas inspects the architecture and reports whether the model is known to fine-tune cleanly. If validation is uncertain, the run still queues but credits are not refunded on training failures for unverified architectures. Two practical bounds: the model must be a text model (there is no path for training on images), and its total parameter count must fit the GPU tier, which realistically means staying under roughly 13B. See Picking a base model for the bring-your-own-model path.

License notes

Every catalogue model's license is summarised here, but you must read the full license before shipping a commercial product. Notable distinctions:

• Apache 2.0 (Qwen 2.5, Qwen 3.5, Mistral 7B): permissive, commercial use allowed without extra conditions.
• Apache 2.0 with a policy layer (Gemma 4): Google relicensed Gemma 4 on 2 April 2026, and the licence text is standard unmodified Apache 2.0, which is a real change from the Gemma Terms of Use that Gemma 3 still carries. Google continues to publish a Prohibited Use Policy alongside it, worded to cover "Gemma or Model Derivatives" without naming a version. Whether that binds under an Apache 2.0 grant is unsettled, so treat Gemma 4 as permissive and still put the policy in front of a lawyer before shipping at scale.
• MIT (Phi-4 Mini): permissive, commercial use allowed.
• LFM Open License v1.0 (LFM2.5 230M / 350M / 1.2B): Liquid AI's own license. Read it before shipping commercially; it is not one of the standard permissive licenses and its terms differ from Apache 2.0.
• Llama Community License (Llama 3.1, Llama 3.2): commercial use allowed up to a monthly active user threshold; products over the threshold need a separate Meta agreement. The threshold is documented in Meta's license text and changes occasionally.
• Gemma Terms of Use (Gemma 3 only): commercial use allowed with specific prohibited-use restrictions Google maintains separately. Read both the Terms of Use and the Prohibited Use Policy. Gemma 4 is Apache 2.0, see above.

Ertas does not enforce license compliance; the model picker lists the license on each card and the models index carries the full text excerpt per model.

What's next