Solar Mini 4: Upstage's 3B-Active Model With a 512K Window

Small models are usually sold on speed and price, then quietly limited on context. Solar Mini 4 does not follow that pattern. Upstage released it on 22 September 2026 with the same 512K-token window and 128K maximum output as its flagship, Solar Pro 4, at a third of the list price.

That makes it tempting to treat Mini 4 as Pro 4 at a discount. It is not, and working out where the two diverge is most of the job of deploying it well.

The shape: 35B total, 3B active

Upstage's model page gives the size plainly: 35B total parameters with 3B active per token, which makes it a mixture-of-experts model. The training data cut-off is February 2026, and the supported languages are Korean, English and Japanese for both input and output, with Upstage putting its Korean fluency first.

Three billion active parameters is a very light forward pass. That is what keeps the price low and the throughput high, and it is also the honest ceiling on how much reasoning happens per token. The 35B figure governs how much memory a deployment needs; the 3B figure governs compute per token. Neither alone predicts quality.

Upstage lists chat, reasoning, structured outputs and tool calling, including parallel tool calls on the catalogue listing. The weights are not public: Upstage's Hugging Face organisation carries tokenizers for its Solar Pro line but not this model, and its recent open-weight releases use the separate Solar Open name. Mini 4 is API-only, with on-premises deployment available by arrangement.

What 512K means on a 3B-active model

A large window and a small active parameter count is an unusual pairing, and it deserves scepticism rather than excitement. Fitting half a million tokens into the prompt is a serving feature. Using information spread across that prompt well is a model capability, and small models tend to lose track sooner.

The window is most useful for jobs where the answer sits in one place and the model only has to find it: searching a long contract for a clause, pulling fields out of a large log dump, answering a question against a manual. It is weakest for jobs that need many distant facts combined, such as reconciling figures across a long report.

Test this on your own documents before trusting it. Context length vs effective context covers how to measure the usable part of a window, which is the number that matters here.

Price, and the discount that ends on 22 October

Upstage lists Solar Mini 4 at $0.10 per million input tokens, $0.40 per million output and $0.01 per million cached input. At the time of writing its console advertises 50% off until 22 October (UTC), which is why some catalogues currently show half those figures. Budget on the list price, not the launch offer.

For comparison, Upstage's model page lists Solar Pro 4 at $0.30 in and $1.20 out, so at list prices Mini 4 costs a third as much per call on a long-document workload where input dominates. Its pricing page also schedules a promotional Pro 4 rate of $0.15 in and $0.60 out from 20 October with no end date shown, which narrows the gap to two-thirds while it lasts. The cached-input rate matters too: if you send the same long document with many different questions, the repeated prefix bills at a tenth of the input price.

Reasoning mode is a cost switch

Mini 4 supports a reasoning mode, and the catalogue lists reasoning-effort control. Upstage's page does not document the effort levels, so check the API reference for the values your route accepts.

Treat reasoning as something you turn on per task rather than by default. Reasoning tokens bill as output, and output is four times the input price here. On a classification or extraction call, reasoning often adds cost without changing the answer. On a multi-step tool plan it can be the difference between a working run and a wasted one. The token cost of reasoning models shows how to measure that trade on your own traffic.

Mini 4 or Pro 4

Upstage positions both as models for agentic work, with the same window, output cap, languages and cut-off. Pro 4's parameter count is not published, so the comparison has to be made on behaviour and price rather than on size.

• Pick Mini 4 for high-volume, bounded calls: routing, classification, extraction, short tool calls, and retrieval over long documents where the answer is local.
• Pick Pro 4 when a failed run is expensive: long agent loops, tasks that combine many facts, or code changes you want right the first time. The Solar Pro 4 guide covers it in detail.
• Run both when the split is unclear. Send work to Mini 4 first and escalate failures to Pro 4. Model routing and fallbacks covers the pattern.

Upstage itself treats Mini 4 as a base for narrow products: its Solar Decide beta, which Upstage says returns typed decisions with calibrated probabilities, is served on Solar Mini 4. That is a reasonable hint about where the vendor thinks the small model is strongest.

A test plan for a small model

The useful question is not whether Mini 4 matches Pro 4 overall. It is which of your tasks it handles at no loss. To find out:

1. Take thirty real requests across your task types, including several long-document ones near your largest real prompt size.
2. Run each on Mini 4 with reasoning off, Mini 4 with reasoning on, and Pro 4, with identical prompts.
3. Score correctness against a known answer, and record output tokens per call.
4. Group the results by task type, not overall. You are looking for the task types where Mini 4 matches Pro 4, because those are the ones to move.

If you work in Korean, include Korean prompts even if most of your traffic is English. Language is where Upstage claims an edge, and it is the part of the claim you are best placed to check yourself.