Platform
Private AI infrastructure — from compute to agents.
Target groups
For enterprise, SMBs, and individual developers.
Knowledge & Support
Everything you need to succeed with Mycelis.
Changelog
All releases, features, and fixes — chronological, transparent, and with context.
Smart Routing — rule-based routing based on token budget, latency, and model type. One VirtualModel slug with multiple deployments behind it.
MCP Library — support for the Model Context Protocol. Expose external APIs as tools directly within the model context.
Three new GPU types: NVIDIA A40, L4, and H100 SXM — for training, inference, and batch workloads.
VirtualModel system — one slug, multiple deployments. Automatic fallback on instance failure.
Routing engine fully revised — 40% lower P99 latency in multi-deployment setups.
API Gateway: streaming responses are now fully SSE-compatible, including long responses.
Token counting for Anthropic models was about ~12% too high — corrected to official cl100k tokenization.
Fine-Tuning Wizard — LoRA training directly in the dashboard. JSONL upload, hyperparameter configuration, and training job launch.
Knowledge Base v2 — multi-document RAG with semantic search. Supports PDF, TXT, Markdown, and DOCX.
BYOK for Google Gemini — route and log your own Gemini API keys through Mycelis.
Workspace roles — admin and member with granular access permissions. Team invitations via email.
Dashboard load time reduced by about ~60% through lazy loading of the deployment list.
Usage dashboard: token consumption can now be filtered by model, deployment, and time range.
Rare race condition when starting multiple training jobs simultaneously.
BYOK OpenAI: timeout for very long completions (>30s) incorrectly failed.
GPU instance deployment — deploy open-source models on dedicated RunPod hardware. Supported models: Llama 3.1, Mistral 7B, Mixtral 8x7B.
OpenAI-compatible proxy endpoint — drop-in replacement for the OpenAI API. Just switch base_url and api_key.
Knowledge Bases v1 — upload documents and use them as context in completions.
PAT-based authentication — Personal Access Tokens for secure API communication.
Managed Keys — OpenAI and Anthropic via Mycelis pay-per-token without your own API keys.
Usage & Billing Dashboard — real-time usage overview with daily billing.