Base Models
The catalog of open-source language models available to finetune on the AI Platform.
The AI Platform's base-model catalog is deliberately narrow for this release: small, permissively-licensed, LoRA/QLoRA-friendly models that finetune in a reasonable time without GPU infrastructure. Larger and GPU-backed models are expected to join the catalog over time.
GET /v2/ai-platform/base-modelsCatalog
| Model | Slug | Family | Checkpoint | Params | License | Cost | Est. Time | Recommended For |
|---|---|---|---|---|---|---|---|---|
| Qwen3.5 0.8B | qwen3.5-0.8b | Qwen | Multimodal | 0.8B | Apache 2.0 | 100 credits | ~25 min | Fastest option — single-label classification |
| Qwen3.5 2B | qwen3.5-2b | Qwen | Multimodal | 2B | Apache 2.0 | 150 credits | ~40 min | Multilingual classification |
| Qwen3.5 4B | qwen3.5-4b | Qwen | Multimodal | 4B | Apache 2.0 | 200 credits | ~70 min | Multilingual instruction-following, entity extraction, agentic tool use |
| Ministral 3 3B | ministral-3-3b | Ministral | Multimodal | 3B | Apache 2.0 | 200 credits | ~65 min | Edge-optimized general-purpose instruction-following with function calling |
| Granite 4.1 3B | granite4.1-3b | Granite | Text-only | 3B | Apache 2.0 | 200 credits | ~60 min | Enterprise RAG, tool-calling, structured JSON output |
| Gemma 4 E2B | gemma4-e2b | Gemma | Multimodal | 2.3B | Gemma Terms of Use | 150 credits | ~50 min | Lightweight reasoning and structured-output extraction |
Cost is a flat, upfront credit charge per run — known before you start, not metered by training time or tokens. See Finetuning Runs for how a run is billed and executed.
Checkpoint indicates the upstream model's native architecture, not what you can train on. Ministral 3, Gemma 4, and Qwen3.5 are natively multimodal (text, image, and — for Gemma 4 — audio) at every size available upstream; Granite 4.1 is a plain text-only model. Regardless of which you pick, the AI Platform only ever finetunes on the text training data you provide — LoRA adapters are scoped to the language-model component only, never the vision/audio encoders, so there's no difference in what data you need to supply.
Training Method
Every base model is finetuned via LoRA (Low-Rank Adaptation), a parameter-efficient finetuning method that trains a small adapter on top of the frozen base model rather than updating all of its weights. This keeps runs fast and affordable on CPU-only infrastructure in this release. The resulting adapter is merged into the base model and converted to GGUF for serving.
Serving
Once a run completes, the merged model is converted to GGUF and registered with Crowdee's Ollama runtime under an organization-scoped tag (crowdee-{organizationId}-{baseModelSlug}), visible in your model registry.
How is this guide?