Solar Mini 4: Upstage's 3B-Active Agent Model, and the Silent Model It Shipped With

Upstage released two things on 22 September 2026, and the second one is the reason to keep reading past the first. Solar Mini 4 is the model: a 35-billion-parameter mixture-of-experts checkpoint with 3 billion parameters active per token, a 512,000-token context window, up to 128,000 tokens of output, and a list price of $0.10 per million input tokens and $0.40 per million output. It went live on the Upstage Console, in the Playground, and as an on-premises deployment, with no preview label and no waitlist. That is the launch, and it is a sensible one. The other thing Upstage shipped the same day was Solar Jev, a decision endpoint served on top of Solar Mini 4 that generates no prose at all — it returns a choice, a score, or a yes/no answer with a probability attached, in a single forward pass, and Upstage bills its output tokens at zero. One of these two announcements is a model release. The other is a small bet that a useful class of agent work does not want a language model to talk.

What actually landed on 22 September

The changelog entry for Solar Mini 4 is short and does not hedge. New model, now available. Upstage's own framing is cost-efficiency for agentic work "where response speed and cost matter" — which is the sentence a lab writes when it wants you to compare it to something three times the price rather than to the frontier.

The spec sheet, per Upstage's model history page:

• Parameters — 35B total, 3B active per token. A sparse mixture-of-experts, so you pay to store 35B and compute with 3B. Upstage does not break out how many experts or how they are routed.

• Context — 512,000 tokens, with a maximum of 128,000 output tokens. Both figures are identical to the flagship's, which is the most interesting line on the sheet.

• Training cut-off — February 2026. Seven months stale at release, which is normal for a model this size and worth knowing before you ask it about anything recent.

• Languages — English, Korean and Japanese. Upstage describes it as "fluent Korean with strong English and Japanese capabilities," which is the correct order for a Korean lab and probably the correct order for the model.

• Capabilities — chat, reasoning, structured outputs and tool calling. Reasoning is a change from the previous generation: the older solar-mini models ignore a reasoning_effort parameter entirely, and Solar Mini 4 does not.

• Model IDs — solar-mini4 is the alias, and it currently resolves to the dated snapshot solar-mini4-260922. A solar-mini4-preview identifier has also been circulating in third-party agent tooling, which is worth knowing if you are writing routing config against a preview string that may not survive.

• Availability — Upstage Console, Playground, and on-premises. No open weights. Upstage has published open weights before (Solar Open 2 went out under a Hugging Face release and its API serving ended on 12 August 2026), but nothing about Solar Mini 4 suggests that is the plan here.

Solar Jev is the part nobody is going to write about

Solar Jev is in beta on a new POST /v1/systemone endpoint. You send a piece of state — a document, a log line, a support ticket — plus a list of typed questions, and you get back a choice from a set you supplied, a score on a rubric, or a yes/no answer, each carrying a calibrated probability taken directly from the model rather than written out as text.

The design consequence is that there is no prose. No explanation, no chain of thought, no preamble. Upstage says each decision is a single forward pass and that output tokens are free. With a 512K context window, the entire state — a whole contract, a whole transcript — can be the input.

This is a narrower product than it first appears, and the narrowness is the point. A model that writes nothing cannot write something wrong. If your pipeline currently asks a chat model to return a JSON verdict and then parses it, and you have spent any time at all defending against the model wrapping that JSON in a friendly sentence, Solar Jev is aimed squarely at you. The trade is that you get a probability and no reasoning trace, so when it is wrong you have nothing to read.

It is worth being precise about what Solar Jev is not: it is not a smaller Solar Mini 4, and it is not a fine-tune you can download. It is a served endpoint on top of Solar Mini 4, available on the Upstage Console, in beta. Beta is the operative word.

What it costs, and the clock on the discount

List price is $0.10 per million input tokens, $0.01 per million cached input tokens, and $0.40 per million output tokens. All figures exclude 10% VAT.

Those list prices are not what you will pay today. Upstage is running a 50% launch discount that its own pricing page dates to 22 October 2026 (UTC) — the underlying structured data on that page gives the window as 22 September 05:00 UTC through 23 October 00:00 UTC, which is the same thing said two ways. During that window the rates are $0.05 input, $0.005 cached, $0.20 output.

Two things follow. First, if you are costing this against anything else, cost it at both numbers, because a 50% discount that expires in a month is not a price, it is a promotion. Second, for context on where the list price sits: the previous Solar Mini generation lists at $0.15 per million tokens in and out with a 32,768-token context and no reasoning mode at all. Solar Mini 4 is cheaper on input, more expensive on output, and gives you sixteen times the context window and a reasoning mode. That is not a refresh. That is a different product wearing the same name.

What nobody has measured yet

Solar Mini 4 is one day old and has no independent evaluation. No Artificial Analysis Intelligence Index, no Agent Arena placement, no third-party agentic benchmark. Upstage has published no benchmark table for it either — the changelog entry carries specs and no scores. Anyone quoting you a number for this model is quoting something else.

There is exactly one third-party-ish data point in circulation, and it needs labelling carefully because it is neither of the two things it looks like. A community repository, hunkim/solar-mini4-jev, wraps Solar Mini 4 behind a compatible shape for the TypeSafe Jev decision API and publishes its own test400 run, judged by another model under a stated rubric. In that run, Solar Mini 4 at reasoning_effort=none scored 98.4% field accuracy against the wrapper's own 94.2%, missing 7 of 447 scored fields against 26; it led on both Korean (98.8%) and English (98.4%); and it averaged 1.21 seconds per call against the wrapper's 0.38 seconds.

Read that for what it is. It is not a vendor benchmark — Upstage did not publish it. It is not an independent evaluation — no neutral body ran it, the harness is the competitor's own, and the judge is a model. It is one developer's self-reported comparison of a hosted API against their own wrapper, and the honest summary is that it is the only signal available and it should not move a production decision on its own. It does suggest two things worth testing yourself: that the model is fast enough for interactive use, and that its Korean is at least as strong as its English.

How to try it without betting a pipeline on it

The awkward part of a day-old model with no independent scores is that the only way to learn whether it fits is to route real traffic through it, and the only way to do that safely is to have somewhere for that traffic to go when it does not.

Failover is the honest answer here, and it is what a routing layer is for. You put Solar Mini 4 behind an endpoint, keep the model you are already running as the fallback, and let the router move traffic when the new one errors, times out or starts returning structured output your parser rejects. That way the first week of a new model is an experiment with a floor under it rather than a launch you have to defend.

One correction worth stating plainly, because it is easy to imply otherwise: OrcaRouter does not route any Upstage model today. Solar Mini 4, Solar Pro 4 and Solar Jev are available through Upstage's own API and several third-party platforms, and if you want Solar Mini 4 specifically, that is where you get it. What we offer is one API key across the 190-odd models we do route, which is useful if the experiment you actually want to run is "does a compact agent model beat what I have" rather than "does this specific Korean model beat what I have." Those are different questions and only one of them has a wrong answer.

The open question

Upstage has now shipped a compact model and a flagship on the same 512K/128K spec sheet, three weeks apart in pricing tiers, and it has put a decision endpoint on the small one rather than the large one. That ordering is deliberate: Solar Jev is served on Solar Mini 4 because a decision that generates no prose does not need a flagship's reasoning, and because you want the cheap model when you are making a million of them.

What is not yet clear is whether 3B active parameters is enough reasoning for agentic work that involves more than extraction. Upstage's marketing says agentic. Upstage's benchmarks say nothing. Until an independent body runs it — and given the release date, that is weeks away at best — the correct posture is to test it on your own task, at the promotional price, with a fallback configured, and to treat any claim about its agentic quality, including the vendor's, as a hypothesis.