Solar Mini 4 is Upstage's new compact AI model built for high-volume agent work: a 35B-parameter mixture-of-experts model that only uses 3B parameters per token, with a 512K-token context window, cheap API pricing, and a free two-week window inside Hermes Agent. If you run AI agents that do lots of repetitive jobs, this one is worth a look.
📺 Watch: Solar Mini 4: FREE Hermes Agent Model!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
I covered it on my channel the day after the free Hermes window opened. Below is what Upstage actually released, the specs and benchmarks they report, what it costs, where you can use it, and my honest take on what it's good for and where it falls short.
What Upstage released with Solar Mini 4
Upstage published its launch post on 1 October 2026. Per Upstage, Solar Mini 4 is its compact model for high-volume agent work. That's the whole pitch. It's not trying to be the smartest model on the planet. It's trying to be the model you can throw thousands of small agent jobs at without burning money.
Per Upstage, it was built for:
- Information extraction (pulling fields out of messy text and documents)
- Structured output (clean JSON-style responses your workflows can parse)
- Tool calling (the thing agents do all day)
- Document validation
- Repetitive agent workflows
Upstage also says it's strong in Korean, plus English and Japanese. So if you run agents for Korean or Japanese markets, that's a real plus.
If you want to learn how to actually put models like this to work inside AI agents, the AI Profit Boardroom has full Hermes and AI agent courses plus Agent OS, for $69/mo.
Solar Mini 4 specs
Here's what Upstage lists in its launch post and docs.
Size. 35B total parameters in a mixture-of-experts design, with 3B active per token. That's why it's fast and cheap. The model has a lot of knowledge stored across its experts, but each token only wakes up a small slice of it.
Context. 512K tokens in, and up to 128K tokens out. That's a big input window for a small model. You can feed it long documents, long chat histories, or big batches of records in one go. The 128K output limit also matters for agents that need to write long structured results.
Speed. Upstage claims 70+ tokens per second sustained at 32 concurrent requests on two H100s. That's Upstage's own number for its own setup, so treat it as a vendor claim. But it lines up with what I saw: it replied fast.
Solar Mini 4 benchmarks (per Upstage)
All of these scores come from Upstage. I haven't run independent benchmarks on them, so read them as the company's reported numbers.
| Benchmark | Score Upstage reports | What it roughly tests |
|---|---|---|
| Artificial Analysis Intelligence Index | 24.1 | Overall intelligence composite |
| AA-LCR | 83.3% | Long-context reasoning |
| SciCode | 47.6% | Scientific coding |
| Humanity's Last Exam | 25.8% | Very hard expert questions |
| AutomationBench-AA | 22.3% | Automation tasks |
| tau3-Banking | 47.2 | Agent tool use in a banking scenario |
Per Upstage, the 24.1 on the Artificial Analysis Intelligence Index is the highest among 3B-active models. That's the key framing. It's not claiming to beat the big frontier models. It's claiming to be the best in its weight class.
The long-context score is the one that stands out to me. 83.3% on AA-LCR, paired with a 512K window, suggests it can actually use that long context rather than just accept it. If your agent reads long documents and pulls things out of them, that's the job this model was made for.
Benchmarks only tell you so much, though. What matters is how a model handles your real tasks. That's why I built Goldie Bench, to test models on the kind of work people actually give their agents.
📺 Watch: Hermes + North Mini Code: New FREE API!
Solar Mini 4 pricing and the launch discount
Per Upstage, the API pricing is:
| Token type | Price per million tokens |
|---|---|
| Input | $0.10 |
| Cached input | $0.01 |
| Output | $0.40 |
On top of that, Upstage is running a 70% launch discount through 10 October (UTC). So if you want to test it on the paid API, now is the cheapest it will be.
Even at full price, those numbers are low. The $0.01 cached input price is the interesting one for agents. Agents send the same system prompt and tool list over and over. If those get cached, your repeat costs drop a lot.
Where you can use Solar Mini 4
Per Upstage's post, you've got five routes:
- Upstage Console API. Direct from Upstage, using the pricing above.
- Solar Chat. Upstage's own chat interface if you just want to try it.
- OpenRouter. Listed as upstage/solar-mini4. Handy if you already route your agents through OpenRouter.
- On-prem. Upstage lists on-prem deployment for companies that need it.
- Hermes Agent through Nous Portal, free for two weeks. Upstage and Nous Research made it free in Hermes starting 5 October 2026.
That last one is the reason most people are searching for it right now. You get a fast, capable agent model inside Hermes for nothing, for a limited window.
📺 Watch: North Mini Code is INSANE (FREE + Local + Open Source)!
How to use Solar Mini 4 in Hermes (short version)
I showed four ways to set it up: Hermes Cloud, Hermes Desktop, Agent OS, and the hermes dashboard command. The quickest is Hermes Cloud: go to portal.nousresearch.com/cloud, create an agent, open the model section, scroll down to "Free", and pick Upstage Solar Mini 4.
The one thing you must get right on every route: pick the FREE Solar Mini 4 variant under Nous Portal, not the paid one. They sit right next to each other.
I wrote up every route step by step in my Solar Mini 4 Hermes setup guide, including what to do when you hit rate limits or can't see the model. If you're running Hermes through Agent OS, my Agent OS guide covers the basics of adding profiles.
What Solar Mini 4 is good at
Here's what I saw when I tested it.
It's fast. Replies came back quickly. I asked it to research the last seven days of AI automation news, and it came back with a sourced breakdown fast. For an agent model, speed matters, because agents make lots of calls per task. A slow model makes the whole agent feel slow.
It handles basic agent tasks fine. Research, summaries, pulling info together, running simple workflows. That's exactly what Upstage built it for, and it does that job.
It's cheap or free. Free in Hermes for two weeks, and low-cost on the API after that. For high-volume work, cost per task adds up fast. A model like this keeps it low.
Long context. 512K in is a lot of room. Per Upstage's long-context score, it can make use of it.
What Solar Mini 4 is not good at
I'll be straight with you. Solar Mini 4 is not frontier level. It's not a replacement for Claude Opus 5.5. If you need deep reasoning, complex multi-step coding, or careful strategic writing, use a bigger model.
Look at the Humanity's Last Exam score Upstage reports: 25.8%. That's solid for a 3B-active model, but it tells you this isn't a model for the hardest problems.
The smart way to use it is as a workhorse. Give your heavy thinking to a frontier model and hand Solar Mini 4 the repetitive stuff: extraction, formatting, validation, quick research, routine agent loops. That's how you get the most out of it without expecting it to be something it's not.
If you want help working out which models should handle which jobs in your business, book a free AI strategy session and we'll map it out with you.
How Solar Mini 4 compares
Here's a quick spec summary so you can see it all in one place. Every number here comes from Upstage.
| Spec | Solar Mini 4 |
|---|---|
| Developer | Upstage |
| Launch | 1 October 2026 |
| Architecture | Mixture-of-experts |
| Total parameters | 35B |
| Active parameters per token | 3B |
| Context in | 512K tokens |
| Max output | 128K tokens |
| AA Intelligence Index | 24.1 |
| Languages Upstage highlights | Korean, English, Japanese |
| API price (in / cached / out) | $0.10 / $0.01 / $0.40 per million |
| Launch discount | 70% through 10 October (UTC) |
| Free access | Hermes Agent via Nous Portal, two weeks from 5 October 2026 |
If you're weighing it against other options for your agent, I keep a running list of the best Hermes Agent models, and a separate breakdown of the best LLM for Hermes Agent depending on what you need it to do.
Should you try it?
Yes, while it's free. There's no downside to testing a fast agent model at zero cost. Set it up in a separate Hermes profile, give it the kind of repetitive jobs your agent does every day, and see how it holds up. If it handles them well, you've found a cheap workhorse for after the free window closes.
Just don't expect it to replace your best model. Use it where speed and cost matter more than raw brainpower.
If you want step-by-step training on building AI agents that use the right model for each job, join the AI Profit Boardroom. You get the Hermes and AI agent courses, Agent OS, and a community of people building this stuff every day.
FAQ
Is Solar Mini 4 free?
Yes, for a limited time inside Hermes Agent. Per Upstage and Nous Research, Solar Mini 4 is free in Hermes through Nous Portal for two weeks starting 5 October 2026. Outside Hermes, the Upstage API costs $0.10 per million input tokens, $0.01 cached and $0.40 output, with a 70% launch discount through 10 October (UTC).
How big is Solar Mini 4?
Per Upstage, it's a 35B-parameter mixture-of-experts model with 3B parameters active per token. It takes up to 512K tokens of input and can write up to 128K tokens of output.
Is Solar Mini 4 open source?
Upstage's launch post doesn't list open weights. The access routes it gives are the Upstage Console API, Solar Chat, OpenRouter, on-prem deployment, and Hermes Agent through Nous Portal. If you need open weights you can download yourself, check Upstage's official channels before assuming they exist.
Is Solar Mini 4 good enough to replace Claude?
No. It's not frontier level and it's not a Claude Opus 5.5 replacement. It's fast and fine for basic agent tasks, which makes it a good workhorse model next to a bigger one.
Want a plan for using Solar Mini 4 and other models in your own agents? Book a free AI strategy session.
