Solar Mini 4: The Specs, the Licence Question, and How to Call It

Solar Mini 4 is Upstage's compact agent model, and its headline trick is arithmetic rather than magic: it carries 35 billion parameters in memory but switches on roughly 3 billion of them for any given token. That is the whole reason Upstage positions it for high-volume, repetitive agent work instead of for the single hardest question you can pose to a model. Upstage released it on 22 September 2026 under the snapshot identifier solar-mini4-260922, served behind the alias solar-mini4, and it sits below the larger Solar Pro 4 in the same Solar family rather than beside it — the family comparison is a separate question and this page is not it. What follows is the standing reference: what the model is, what it costs, whether you can download it, what Upstage claims, what an independent evaluator actually measured, and how you call it today.

Every number below carries its source inline, and the sources are of two kinds. Figures attributed to Upstage are the vendor's own published claims, read from Upstage's model documentation and model-history pages and from Upstage's own 1 October 2026 announcement post. Figures attributed to Artificial Analysis are that organisation's independent measurement, published 30 September 2026. Where the two disagree, both are printed rather than reconciled.

What Solar Mini 4 is, and who publishes it

Upstage is the publisher. The model's documentation page describes it in one line as a "cost-efficient compact model for agentic use", and that framing is doing real work: the model is aimed at workloads where an agent calls a model many times in a row — reading a form, deciding what to do next, calling a tool, writing a structured result — rather than at workloads where a single answer has to be as strong as possible.

The identifiers matter more than they look, because Solar Mini is a name Upstage has reused for years:

• The alias you pass as the model name is solar-mini4.

• The dated snapshot behind that alias is solar-mini4-260922, and the release date Upstage publishes for it is 22 September 2026.

• Every earlier Solar Mini snapshot on Upstage's model-history page — the 2024 and 2025 releases, including the Japanese variants — carries a deprecated marker and a migration notice pointing at the solar-mini4 alias. If you are holding an integration written against an older Solar Mini, the alias you were told to move to now resolves to this model, not to a 10.7B one.

• The model-history page also lists Solar Mini 4 as the base model behind Upstage's separate solar-decide beta endpoint. That endpoint is a different product with its own pricing and behaviour; it is not covered here.

Size and architecture

Upstage's own model page states the architecture explicitly: 35B total parameters with 3B active. This is a sparse mixture-of-experts design, meaning the full weight set has to exist in memory even though only a slice of it is used per token.

Two consequences follow, and both are the vendor's own:

• The memory requirement is set by the 35B figure, not the 3B figure. Upstage's announcement post says plainly that "the full 35B model weights must still be loaded into memory", and adds that a quantized form can be run on a single 80GB H100 for on-premises deployment. The two-H100 configuration used in the vendor's throughput test is described as a different configuration from the minimum deployment requirement.

• The comparison class is other 3B-active models, not other 35B models. Upstage compares Solar Mini 4 against Qwen3.6 35B A3B (also 3B active) and Nemotron 3.5 Lightning in its announcement, and the independent ranking places it against that class too. Solar Pro 4 is a larger sibling in the same family and a different purchase; nothing on this page should be read as a Pro 4 comparison.

Are the weights open? No licence is published

This is the question the model's own name invites, and the honest answer as of 13 October 2026 is that Upstage publishes no licence for Solar Mini 4 and no open-weights statement for it, on either its model page or its model-history page. There is no licence field, no named licence, and no weights download offered for this model.

The absence is more legible when you set it against Upstage's own behaviour elsewhere, because Upstage does use open-weights language when it means it:

• Upstage's model-history page describes solar-open2 as remaining "available as open weights on Hugging Face" and calls it Upstage's open-weight model.

• The same page calls solar-10.7b-v1.0 "the initial, open-source version of Solar Mini".

• In Upstage's public Hugging Face organisation you can find tokenizers for several Solar models and full weight repositories for others, but no Solar Mini 4 checkpoint.

So the contrast is deliberate and visible: the vendor labels the Solar models it has opened, and Solar Mini 4 is not one of them. What Upstage offers instead is an on-premises deployment path — the console's platform list names "On-premises" alongside Upstage Console and the Playground, and the announcement post notes the quantized model fits one 80GB H100. On-premises deployment behind a vendor agreement is not the same thing as open weights, and this page will not blur the two.

Context window and maximum output

Upstage's model page states 512K tokens of context and up to 128K output tokens. The model-history page repeats both figures for the solar-mini4-260922 snapshot.

One number here does not agree across sources, and it is the kind of disagreement worth printing rather than smoothing. Artificial Analysis lists the context window as roughly 1.05M tokens — about double what Upstage publishes. The vendor's own figure is the one to build against, because the vendor controls the endpoint; the independent listing is the reason to test your actual long-document case rather than trusting either number. If you are sizing a workload around a very large context, treat 512K as the contract and 1.05M as an unverified claim you would have to confirm against the API yourself.

Modalities, and where Korean sits

Solar Mini 4 is a text-in, text-out model, and that is not an assumption drawn from its name — it is what the vendor's capability listing shows. The model page lists chat and structured outputs and no other input or output type: there is no image input, no audio input, no file or video modality offered for this model.

Language coverage is the part of the specification most likely to matter to the reader, and Upstage is unusually specific about it. The model page lists three languages — English, Korean and Japanese — and marks each of them "Input and output", meaning they are listed as production languages in both directions rather than as input-only comprehension. Upstage's own description of the model makes the same point in prose: it advertises fluent Korean alongside strong English and Japanese support.

For a Korean-language workload, that is a first-class claim rather than a transliteration of an English-first model, and it is a vendor claim — Upstage is a Korean company and its benchmarks for this model are run in Korean, including the cost test described below. No independent evaluator has published a Korean-language evaluation of Solar Mini 4 that we could find, so the Korean-quality claim is the vendor's alone.

What Upstage claims, and what was independently measured

Upstage's 1 October 2026 announcement post reports: an Artificial Analysis Intelligence Index v4.3.2 score of 24.1, which it says is the highest overall score among the 3B-active models in that comparison; 83.3% on AA-LCR, a long-context reasoning evaluation; 47.6% on SciCode; 25.8% on Humanity's Last Exam; 22.3% on AutomationBench-AA, which covers multi-step workflows across real SaaS applications; and 47.2 on τ³-Banking.

Those are vendor-reported figures. The same post describes the throughput test behind its speed claim: more than 70 tokens per second per request while processing 32 concurrent requests on two H100 GPUs, measured with 4K input tokens and 1K output tokens, without prefix caching, over a load test lasting more than 600 seconds. The conditions are the vendor's and the result is the vendor's.

Artificial Analysis, publishing its own evaluation on 30 September 2026, measured the model on the same suite the vendor cites and put the Intelligence Index at 24.05 on v4.3.2 — the same number to rounding — with a cost of $0.3634 per index task, an average of 88,300 output tokens per task of which 71,600 were reasoning tokens, a throughput of 204.07 tokens per second, and roughly 430 seconds per task. Artificial Analysis's figures are its own measurement, produced by running the model through its harness; they are not a summary of the vendor's claims, and the cost-per-task figure in particular is one Upstage does not publish.

Three honest limits belong with those numbers:

• The vendor publishes no independent head-to-head at matched effort. Upstage's cost comparison against MiMo-V2.5 — $1.25 versus $2.20 to complete 1,000 three-task sets, and $0.36 versus $0.61 for 1,000 runs of a multi-step tool-calling task — is described by Upstage as an initial internal test, run on three Korean-language tasks with three runs each at temperature 0 and no prompt caching, with the comparison model reached through a third-party router and latency excluded. Two models were run in that test, but not under an independently controlled harness, so the result is a vendor estimate of a vendor-favourable setup rather than a benchmark.

• There is one thing the vendor does not publish at all. Upstage reports no Terminal-Bench result and no long-form agentic-coding score of its own for Solar Mini 4. If agentic coding is your use case, the vendor gives you no number to reason with.

• Reasoning is off by default. Upstage states that Solar Mini 4 does not use reasoning unless you ask for it, and that developers can set a reasoning-effort parameter to match task complexity. That single setting will move every token-count and cost figure above, in either direction, so any comparison that does not state its reasoning setting is not comparable to another one.

What it costs, and how to call it

Upstage's model page lists standard rates of $0.10 per million input tokens, $0.01 per million cached input tokens, and $0.40 per million output tokens.

The promotional pricing around those rates does not agree with itself, and it is worth saying so plainly. The banner on Upstage's console pages reads "Solar Mini 4 is live. 50% off through Oct 22 (UTC)". The announcement post, dated 1 October 2026, says something different and internally impossible: "Upstage Console offer a 70% launch discount through October 10 UTC, with standard pricing taking effect on October 1 UTC" — a discount that ends before the date it takes effect, and one that no longer matches the banner on the vendor's own site. Upstage also notes that promotion dates and terms are subject to change. The reliable current statements are the console banner and the standard rates on the model page; treat any other discount figure you find, including in the vendor's own post, as superseded.

To call the model, Upstage's getting-started documentation shows the vendor's OpenAI-compatible client pointed at an Upstage base URL, with the key created from the Upstage Console:

• Base URL: https://api.upstage.ai/v1, with chat completions called through that client's chat interface and the model set to solar-mini4. The documentation's own example uses a different Solar model name in its sample call, so substitute the alias rather than copying the snippet verbatim.

• An optional reasoning-effort parameter sits alongside the model parameter in that same example, which is the vendor's documented way of turning the default-off reasoning on.

• The model supports tool calling and parallel tool calling, JSON mode and JSON-Schema-based structured outputs, per the model page and the announcement post.

• Upstage's model page states that the API collects no data from this endpoint: input data is not used for training and is not stored beyond what service delivery requires.

• There is no published rate limit for this model. Upstage's documentation does not list requests-per-minute or tokens-per-minute figures for Solar Mini 4 anywhere we could find, so any throughput ceiling you hit is one you will have to discover empirically rather than read.

• Besides the API, the model is reachable through Upstage's own Console and Playground — the model page notes the Playground can be tried without an API key — and through an on-premises deployment that Upstage handles by contact rather than by download.

One practical note on pricing arithmetic across providers: OrcaRouter charges 0% markup and passes provider list prices through unchanged, which means a vendor rate change lands in the router's numbers the same day it lands in the vendor's. That matters here specifically because the discount above has a published end date — after it, the effective rate for this model moves without any announcement, and a pass-through router moves with it instead of re-pricing on a slower cycle.

The short version

Solar Mini 4 is a 35B-parameter, 3B-active sparse mixture-of-experts text model from Upstage, released 22 September 2026, with a 512K-token context window, 128K maximum output, English/Korean/Japanese support in both directions, tool calling and structured output, and a price of $0.10 per million input tokens and $0.40 per million output tokens. Its licence status is the open question, and the answer today is that there isn't one: Upstage publishes no licence and no weights for this model, in contrast to the Solar models it does label as open weights. Its speed and cost profile is genuinely well supported — the vendor and an independent evaluator agree on the headline index score to rounding — while its Korean-language quality and its discount terms rest on vendor statements alone. If you need the model's weights, this is not the model. If you need a cheap, fast, Korean-capable agent endpoint behind an API, it is, and you can be calling it within a few minutes of creating a key.