Solar Mini 4 is Upstage’s new agent model — 35 billion parameters, only 3 billion active, a 512K-token context window, and right now you can use it free inside Hermes Agent. I tested it in the video below on a real research task. Here are the specs, the benchmarks, the price and my honest take.
Key takeaways
- Solar Mini 4 is a 35B mixture-of-experts model with 3B active parameters, from South Korea’s Upstage.
- It has a 512K context window and up to 128K output tokens.
- It scored 24.1 on the Artificial Analysis Intelligence Index.
- Standard API pricing is $0.10/$0.40 per million tokens — and it’s free in Hermes Agent for a limited time.
What Solar Mini 4 is
Upstage is a South Korean AI company, and Solar Mini 4 is a model it pretrained from scratch specifically for agents — high-volume, repetitive work like retrieving information, producing structured output and calling tools, at low cost. Because it’s a mixture-of-experts model, only 3 billion of its 35 billion parameters are used for each token, which is why it’s so fast.
| Solar Mini 4 spec | Detail |
|---|---|
| Developer | Upstage (South Korea) |
| Architecture | Mixture of experts — 35B total parameters, 3B active per token |
| Context window | 512K tokens (Upstage’s figure) |
| Max output | 128K tokens |
| Artificial Analysis Intelligence Index | 24.1 — the top score among 3B-active models in Upstage’s comparison |
| Standard API price | $0.10 per 1M input tokens, $0.40 per 1M output tokens |
| Where to use it | Upstage Console API, Solar Chat, OpenRouter, Hermes Agent (free for a limited time) |
Solar Mini 4 benchmarks
Upstage says Solar Mini 4 scored 24.1 on the Artificial Analysis Intelligence Index — the highest of the 3B-active models it compared against, and above models with around ten times the active parameters. Here are its other published scores:
| Benchmark | Solar Mini 4 | What it measures |
|---|---|---|
| AA-LCR | 83.3% | Long-context reasoning |
| SciCode | 47.6% | Code-based scientific problem solving |
| τ³-Banking | 47.2 | Applying policy and using tools |
| Humanity’s Last Exam | 25.8% | Very hard expert questions |
| AutomationBench-AA | 22.3% | Real SaaS workflow automation |
Always test it on your own tasks — benchmarks published by the model maker are a starting point, not a verdict.
Want this working in your business, not just bookmarked? Using fast free models to run your AI agents is exactly what we build together inside the AI Profit Boardroom — 3,400+ members, four live calls a week, and a full one-hour DeepSeek Harness course in the classroom.
Prefer it mapped 1-on-1 first? Book a free strategy session and we’ll plan it for your exact situation.
Join the AI Profit Boardroom →Book a Free Strategy Session →
How fast Solar Mini 4 is in practice
In my test I asked it to research the last seven days of AI automation news. It came back quickly with a tidy breakdown — sources, what stands out, the key stories. It even surfaced a launch I hadn’t seen. Upstage quotes 70+ tokens per second on two H100 GPUs.
Let me be 100% honest: this is not a frontier model. You won’t use it the way you use Claude Opus 5.5 for hard reasoning. But for basic agentic tasks — research, summaries, structured output — it’s more than good enough, and it’s fast.
Solar Mini 4 pricing and where to use it
Standard pricing is $0.10 per million input tokens, $0.01 per million cached input tokens and $0.40 per million output tokens, and Upstage ran a 70% launch discount until 10 October 2026. You can use it through the Upstage Console API, Solar Chat, OpenRouter, on-premises, and inside Hermes Agent.
The best deal right now: Hermes Agent users get it free for a limited window, which opened on 5 October. My step-by-step Solar Mini 4 Hermes guide shows exactly how to set it up — and the one setting that stops you landing on the paid version by accident.
Who should use Solar Mini 4
- Hermes Agent users who want a fast free model for everyday tasks
- Builders running lots of agent calls who care about cost per task
- Long-document work — the 512K context fits huge inputs
- Korean-language work — Upstage builds with native Korean in mind
If you need deep reasoning or complex coding, pair it with a frontier model and use Solar Mini 4 for the high-volume grunt work.
Solar Mini 4 vs other small agent models
| Model | Maker | Active params | Context | Licence |
|---|---|---|---|---|
| Solar Mini 4 | Upstage | 3B (of 35B) | 512K | Commercial API |
| Kolibri-1 | Aleph Alpha | 3.46B (of 78B) | Up to 1M | Apache 2.0, open weights |
| Ling 3.1 Flash | inclusionAI | Small MoE | Long context | Free in Hermes |
| LFM 2.5 | Liquid AI | Small dense | Shorter | Open weights |
Solar Mini 4’s edge is the combination of speed, a huge context window and a very low price, with strong tool use for its size. If you want open weights you can run yourself, look at Kolibri AI. If you want the fastest thing on a Mac, a tiny local model like LFM 2.5 is hard to beat.
Solar Mini 4 use cases I’d try
- Daily news research — the exact test I ran in the video.
- Summarising long documents — the 512K context swallows big PDFs and transcripts.
- Structured data extraction — JSON mode and structured outputs are built in.
- First-draft content briefs — fast and cheap to iterate.
- Sub-agent work — let a frontier model plan, and Solar Mini 4 do the repetitive steps.
The pattern that works: give the cheap, fast model the high-volume tasks and save the expensive model for the decisions.
Getting the best results from Solar Mini 4
Small, fast models reward clear instructions. A few habits that help:
- Be specific — say exactly what you want back, how long, and in what format.
- Ask for structure — tables, bullet points or JSON play to its strengths.
- Give it the source material — with 512K of context, paste the documents rather than relying on what it knows.
- Check the facts — like any model, it can be confidently wrong, so verify anything important.
- Escalate when needed — if it struggles on a hard problem, switch to a frontier model rather than re-prompting forever.
My verdict on Solar Mini 4
Solar Mini 4 is a fast, cheap, capable agent model, and while it’s free in Hermes there’s no reason not to try it. Just remember it can get rate limited — keep another free model ready to switch to.
I’ll keep this page updated as pricing and availability change.
FAQ: solar mini 4
What is Solar Mini 4?
Upstage’s agent-optimised mixture-of-experts model: 35B total parameters, 3B active, 512K context and up to 128K output tokens.
Is Solar Mini 4 free?
It’s free inside Hermes Agent for a limited time from 5 October 2026. Otherwise API pricing is $0.10 per million input and $0.40 per million output tokens.
How good is Solar Mini 4?
It scored 24.1 on the Artificial Analysis Intelligence Index, top among 3B-active models in Upstage’s comparison — strong for agent tasks, but not frontier-level.
Where can I use Solar Mini 4?
The Upstage Console API, Solar Chat, OpenRouter, on-premises, and Hermes Agent.
Who makes Solar Mini 4?
Upstage, an AI company based in South Korea.
Two ways I can help from here. Join the AI Profit Boardroom — it’s where your AI agent model setup gets built with 3,400+ members doing the same.
Or grab a free strategy session and bring your questions.
Join the AI Profit Boardroom →Book a Free Strategy Session →
About Julian Goldie
I’m Julian Goldie — founder of Goldie Agency, best-selling author, and an AI educator with 428K+ YouTube subscribers. I’ve spent 10+ years in SEO and online business and I test every AI agent on my own work before I recommend it. Join the AI Profit Boardroom for the daily builds, or book a free strategy session to talk through yours.
Related reading
Last updated October 2026. This page is a living guide to solar mini 4 — the facts here move fast and it is updated as they do.
