Ertas AI
Ertas AI provides a visual fine-tuning platform that lets app developers train small, open-source language models on cloud GPUs and export them as GGUF files for on-device deployment. The platform eliminates per-token API costs, network latency, and data-privacy concerns by running models directly on user devices via llama.cpp, Ollama, or LM Studio. A fine-tuned 1.2B model achieved 35.41 ROUGE-1 on QMSum, outperforming five hosted frontier models while running 23× faster and at 99% lower cost than Claude Opus 5.
- Artificial Intelligence
- Developer Tools
- Software Only
- AI Agents
Funding
Founders
Founder details are not available yet.
Product
Problem
Cloud-based AI APIs create scaling problems for mobile app developers: per-token costs grow linearly with user count, network round-trips add latency and fail offline, and user data leaves the device for third-party servers. Developers also face deprecation risk when API providers retire models, leaving apps broken on someone else's timeline. At 10,000 users, cloud API costs can exceed $620/month, while every inference adds 200-800ms of latency and sends user data to external servers.
Solution
Ertas AI provides a visual training canvas that lets app builders fine-tune open-source language models without writing Python or managing infrastructure. Users upload JSONL datasets or import from Hugging Face, select a base model from a curated catalog, and configure training recipes visually. The platform handles the training pipeline on cloud GPUs and offers one-click export to GGUF, the open format compatible with llama.cpp, Ollama, and LM Studio. The resulting models run entirely on-device, delivering instant inference with zero per-token cost, full offline capability, and privacy by architecture since user data never leaves the phone. The platform also includes Data Craft, a guided dataset builder for creating training data from scratch when no dataset exists. Ertas research shows a fine-tuned 1.2B model scoring 35.41 ROUGE-1 on QMSum versus 32.43 for the best of five hosted frontier models, while running 23× faster and at 99% lower cost than Claude Opus 5.
Target Audience
Primary customers are indie developers, startup founders, agencies, and solo builders shipping AI-powered mobile or desktop applications who need custom models without ML engineering resources.
Features
- Visual training canvas with drag-and-drop dataset upload, base model picker, and recipe configuration—no code or ML expertise required
- Curated model catalog spanning Llama, Mistral, Phi, Gemma, Qwen, and LFM families with parameter counts from 230M to 22B+ and license badges for commercial-use decisions
- Bring-your-own-model support via Hugging Face URL validation for models outside the catalog
- One-click GGUF export producing open-format files compatible with llama.cpp, Ollama, and LM Studio, with no proprietary lock-in
- Two GPU tiers—T4 (16 GB VRAM) for models under 5B parameters and A10G (24 GB VRAM) for larger models with longer contexts
- Data Craft guided dataset builder for creating training data from scratch when no dataset exists
- Domain-specific fine-tuning that achieves 94% task accuracy versus 71% with GPT-4 prompt engineering, per the company's benchmarks
- Ertas Deployment CLI for installing trained models into applications or registering them with Ollama