Supported models

    The base models Ertas can fine-tune today, with licenses, parameter counts, GPU tier, GGUF size, and notes on what each is best for.

    This page is the single-table view of every base model currently in the Ertas catalogue. For the full marketing-grade view (license excerpts, taxonomy, dataset compatibility notes) see the models index. For the size math behind the GGUF column, see File sizes and formats. For why a 5B total-parameter model can need A10G even when the compute-parameter count is lower, see Concepts on Gemma 4 E2B.

    The catalogue

    GGUF sizes are at Q4_K_M (the only quantisation Ertas exports today, see Quantization). "Free plan?" answers whether a Free-plan account can train this model; A10G models require Builder or higher.

    Llama family

    ModelParamsLicenseGGUF size (Q4_K_M)GPU tierFree plan?Best for
    Llama 3.2 1B Instruct1.0BLlama Community License0.81 GBT4YesSmallest official Llama; web and mid-range mobile
    Llama 3.2 3B Instruct3.0BLlama Community License2.02 GBT4YesMobile and desktop sweet spot
    Llama 3.1 8B Instruct8.0BLlama Community License4.92 GBA10GNoDesktop and server-side inference where the extra quality matters

    Mistral family

    ModelParamsLicenseGGUF size (Q4_K_M)GPU tierFree plan?Best for
    Mistral 7B Instruct7.0BApache 2.04.37 GBA10GNoPermissively-licensed general purpose

    Mixtral and other mixture-of-experts variants are not yet trainable in Ertas; see Known limitations.

    Phi family

    ModelParamsLicenseGGUF size (Q4_K_M)GPU tierFree plan?Best for
    Phi-3 mini 4k Instruct3.8BMIT2.39 GBT4YesPermissive license, small footprint, reasonable quality
    Phi-4 Mini Instruct3.84BMIT~2.4 GBT4YesNewer Phi generation; stronger instruction-following than Phi-3 mini at the same class

    Gemma family

    ModelParamsLicenseGGUF size (Q4_K_M)GPU tierFree plan?Best for
    Gemma 3 1B IT1.0BGemma Terms of Use0.81 GBT4YesSmallest Gemma; rough parity with Llama 3.2 1B
    Gemma 3 4B IT4.0BGemma Terms of Use2.49 GBT4YesSlightly larger than Llama 3.2 3B; same class
    Gemma 4 E2B5.1B total / 2.3B effectiveGemma Terms of Use3.19 GBA10GNoHigher quality at 3B-class inference; needs A10G due to embedding tables

    Qwen family

    ModelParamsLicenseGGUF size (Q4_K_M)GPU tierFree plan?Best for
    Qwen 2.5 0.5B Instruct0.5BApache 2.00.49 GBT4YesTiny baseline; useful for testing pipelines, not production
    Qwen 2.5 1.5B Instruct1.5BApache 2.01.12 GBT4YesMultilingual support, permissive license
    Qwen 2.5 3B Instruct3.0BApache 2.02.10 GBT4YesStrong structured-output and multilingual at 3B class
    Qwen 2.5 7B Instruct7.0BApache 2.04.68 GBA10GNoHigh-quality multilingual and long-context
    Qwen 2.5 14B Instruct14.0BApache 2.08.99 GBA10GNoLargest catalogue model; desktop-only deployment

    Code-specialised

    ModelParamsLicenseGGUF size (Q4_K_M)GPU tierFree plan?Best for
    Qwen 2.5 Coder 1.5B1.5BApache 2.01.12 GBT4YesLightweight code completion; FIM support out of the box
    Qwen 2.5 Coder 3B3.0BApache 2.02.10 GBT4YesCode completion recipe default; see Cookbook: code completion
    Qwen 2.5 Coder 7B7.0BApache 2.04.68 GBA10GNoDesktop coding assistant tier

    Under active evaluation (not yet in catalogue)

    Models in this group are being tested and validated for catalogue inclusion. They are not yet selectable in the Studio model picker; the models index is the source of truth for what is currently selectable.

    ModelParamsLicenseStatus
    Llama 4 (when generally available)TBDLlama Community LicenseUnder evaluation
    Phi-4 (full, ~14B)~14BMITUnder evaluation
    Qwen 3 familyTBDApache 2.0Under evaluation
    TinyLlama 1.1B Chat1.1BApache 2.0Under evaluation; useful as a very-low-footprint sub-3B candidate
    SmolLM 1.7B Instruct1.7BApache 2.0Under evaluation; sub-3B with stronger instruction following than TinyLlama

    If you are relying on one of these for an in-flight project, the Hugging Face URL import path lets you train against them as unverified architectures today, at the cost of giving up automatic credit refunds on training failures. Catalogue inclusion will lift the unverified status when validation completes.

    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-3 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, prefer 1B-class (Llama 3.2 1B, Gemma 3 1B, Qwen 2.5 1.5B) to keep first-load downloads under 1 GB and stay inside browser memory ceilings.

    Hugging Face models outside the catalogue

    You can fine-tune any Unsloth-compatible model from a Hugging Face URL. Ertas validates 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. 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, Qwen 2.5 Coder, Mistral 7B): permissive, commercial use allowed without extra conditions.
    • MIT (Phi-3 mini, Phi-4 Mini): permissive, commercial use allowed.
    • 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, Gemma 4): commercial use allowed with specific prohibited-use restrictions Google maintains separately. Read both the Terms of Use and the Prohibited Use Policy.

    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