Playing this video loads it from YouTube (Google), which may store cookies on your device and receive your IP address. Click to play.
Video: "Solar Mini 4: FREE Hermes Agent Model!" by Julian Goldie on YouTube.
What Solar Mini 4 actually is
Solar Mini 4 is a mixture-of-experts model from Upstage, a South Korean AI lab. Total parameter count is 35 billion, but only 3 billion are active at any one time - the sparse MoE design routes each request through the relevant subset, which keeps inference fast without sacrificing much capability. The context window is 512,000 tokens, which is practical for most agent tasks involving long documents or multi-step workflows.
Upstage built it specifically for high-volume agent workloads: structured data extraction, tool calling, and repeated task loops. It is not a conversational model aimed at general chat; it is designed for the kind of mechanical, repetitive work that Hermes Agent runs through all day.
How the free window works
From 5 October 2026, Solar Mini 4 is available free via the Nous Portal inside Hermes. Upstage has not published a firm end date, but the arrangement is described as a limited-time partnership - two weeks is the estimate doing the rounds. If you want to test it at no cost, now is the time.
Outside the free window, the API is priced at $0.10 per million input tokens and $0.40 per million output tokens. Upstage ran a 70% launch discount until 10 October, so if you are reading this after that date, the full rate applies. Still relatively cheap for a 35-billion-parameter model, but not free.
The setup, and the one thing to get right
In Hermes Cloud or Hermes Desktop, you create a new agent profile, open the model list, and search for Solar Mini 4. It appears twice: one entry is the free Nous Portal version, the other is the paid API. Julian flagged this as the main pitfall - the labels are similar enough to pick the wrong one by accident. The free version is marked as a Nous Portal model; the paid API entry connects directly to Upstage's billing.
Once you have selected the correct one, the model loads into the agent like any other. Your existing goals, skills, and memory carry over without any rebuilding.
What it handles well, and what it does not
Julian's tests found Solar Mini 4 fast and reliable at the tasks it is designed for: building skill files, running tool calls, structured content generation, and the kinds of repetitive agent loops that eat through tokens quickly. The low active-parameter count keeps latency down even on longer context windows.
That said, it is not a frontier model. Deep chain-of-thought reasoning and complex multi-agent coordination are better handled by Claude Fable or similar. For anything that involves nuanced judgement rather than mechanical execution, you will notice the gap. The honest use case is as a workhorse model for high-frequency tasks, not as a replacement for your main reasoning model.
Where this connects to NordSys
One of the recurring questions we get from clients running Hermes-based agent setups is about cost - specifically, whether the model subscription fees add up at scale. Models like Solar Mini 4, available free for a limited window, change that calculation for lightweight automation jobs. If you are considering an AI agent deployment and want to understand how model selection affects both capability and cost, our AI Agents service covers exactly that - from initial setup through to day-to-day management.
See our AI Agents service →