--
Models
0
Loaded
--
Deps OK
--
Version
Quick Actions
Deploy
Output
Ready.\n
Period:
--
Developers
--
Active Apps
--
Calls Today
--
Avg Latency
Plan Distribution
API Calls Over Time
Top Models by Calls
Loading…
Top Apps
Loading…
Top Developers
Loading…
ollama pull
mimicxai/darwin
--
Total Pulls
--
This Period
--
Unique IPs
Daily Pulls
No data
Top Regions
--
--
All Time
--
Last 30d
--
Last 7d
Daily Installs
No data
By OS / System
--
Foundation Models
| Family | Task | Architecture | Base Model | Parameters | Status | Action |
|---|
Darwin Model Lineup (shown in the chat model picker)
| Name | Model ID | Description | Backend model | Enabled | Action |
|---|---|---|---|---|---|
| Loading… | |||||
Train a Model
(own your weights — fine-tune an open base)
Fine-tune a permissively-licensed base (Apache-2.0: Qwen, OLMo) on your skill
catalog so tool-calling is reliable and the weights are legally yours to rebrand.
Runs on the GPU host; exports GGUF and registers an Ollama model.
Train Image Model
(SDXL / FLUX LoRA — separate from LLM training)
Fine-tune a diffusion model (SDXL / FLUX) on your own image + caption pairs.
This is a different stack from LLM training — output is a LoRA .safetensors
that loads in ComfyUI / diffusers / Automatic1111 (not Ollama). Data folder:
image files each with a matching .txt caption.
Pull Model
(from Ollama or Hugging Face)
Pull a model so it's available locally. Ollama pulls make a model runnable
for serving; Hugging Face pulls download the full weights folder needed for
merging or fine-tuning.
Merge Models
(combine fine-tunes into one best model)
Combine several fine-tuned models that share a base (e.g. a code model + a
chat model + a skill-calling model, all Qwen2.5) into one model carrying the
strengths of each. Runs on the worker via mergekit — minutes, no GPU training.
Models must share the same architecture (you can't merge an LLM with an image model).
Live Inference
Waiting for input...\n
GPU Providers
Register external GPU inference endpoints (Nebius, ComfyUI, a custom Stable Diffusion server…). The agent dispatches media generation to the active provider of the matching kind.
Enrol a GPU (recommended)
On your GPU box (Nebius VM, etc.) run this one line. Nothing to type or paste — the node will print a short code that you approve under Pending approvals below.
No secret in the command · approval happens here
Prefer a pre-approved token instead? (skips the approval step)
Pending approvals
No nodes waiting.
Add a provider (endpoint)
Quick add by URL
Works for OpenAI-compatible endpoints (vLLM, Nebius, OpenAI, Together…). Paths are inferred.
Advanced / non-standard endpoint
Registered providers
| Name | Kind | URL | Enabled | Billing | |
|---|---|---|---|---|---|
| Loading… | |||||
Preloaded models per node
Each node preloads a model per capability at startup. Change one here and the node swaps it live (unloads the old, re-warms the new) — no reinstall. Loaded models show ✓.
Loading…
Darwin Skills
Register function-calling skills. Compatible with Distribute skill format.
Distribute Built-in Skills
Toggle to enable for Darwin
Loading catalog from Distribute…
AI Models (MaaS)
Toggle on to bind as a skill for Darwin
Loading models…
Custom Skills
| Skill | Backend | Status | Actions |
|---|---|---|---|
| Loading… | |||