Solar Mini 4 splits Upstage into Pro and Mini

Solar Mini 4 is Upstage's compact model for repetitive work, announced in Korean business coverage on October 1, 2026. Chosun Biz timestamped its story at 9:14 that morning. DigitalToday's English item carries a published time of 9:04:56 the same day and an update line of 10:15 a.m. KST on October 2.

TL;DR - Both pages describe a 35-billion-parameter mixture-of-experts model that activates 3 billion parameters during inference. - Upstage's own reading of Artificial Analysis puts an Intelligence Index score at 24. That is a company claim, not a leaderboard screenshot in either article. - Chosun Biz adds a 70 percent API discount through October 10 and names rival models. DigitalToday does not.

What actually happened

Chosun Biz says Upstage unveiled Solar Mini 4 on the 1st and will split its large-language-model lineup into Solar Pro and Solar Mini. Pro is for complex agent tasks that need sophisticated judgment. Mini is for repetitive work such as step-by-step automation, where speed and expense matter more than peak difficulty. The point of the split, as the story tells it, is that a company can combine the two based on task complexity and throughput instead of sending every job to the larger model.

The architecture claim is the same on both pages. Total parameters are 35 billion. During inference the model selectively activates 3 billion of them. Chosun Biz says that low active count is the edge on expense and response speed. With quantization, it can run on a single GPU, which the company presents as a fit for on-premise servers when security rules discourage sending data to an outside cloud. DigitalToday repeats the single-GPU and on-premise point and says the focus was lowering inference cost for repetitive data processing and work automation.

Both pages say the model is offered as an API through OpenRouter. Chosun Biz adds history: in August, Solar Pro 4 was the first Korean model officially listed on OpenRouter and once ranked third weekly by usage of closed-model tokens there. DigitalToday only says Solar Pro 4 was listed on OpenRouter in August. It does not repeat the weekly rank. Where the second page is silent, this article does not treat the rank as confirmed twice.

DigitalToday's lede says the model was unveiled "on Wednesday," while its published timestamp is October 1, 2026. October 1, 2026 was a Thursday. The page does not explain the mismatch. Chosun Biz says "on the 1st," which matches the timestamp. Use the dated line, not the weekday word, until Upstage publishes its own clock time.

What the score of 24 does and does not mean

Chosun Biz says Solar Mini 4 scored an overall 24 on the Intelligence Index, which it abbreviates AAII, from Artificial Analysis. It says the model outperformed peers including Alibaba's QwQ 3.6-35B-A3B and Nvidia's Nemotron 3.5 Lightning. It also says the model matched or exceeded models with larger active parameter counts, including Nvidia's Nemotron 3 Ultra, Google's Gemma 4 31B, and Xiaomi's MiLMo V2.5. Those names appear only on the Chosun Biz page opened for this piece. DigitalToday says the same overall score of 24 and that Upstage called it a high level of performance among models with 3 billion active parameters. It does not list the rivals.

A company saying it "outpaced every competing model in its class" is a quote, not a table. Kim Seong-hoon, identified as Upstage's chief executive, told Chosun Biz that Solar Mini 4 set a new standard for compact models and that the Pro and Mini split would let companies choose. DigitalToday quotes Sung-hoon Kim, the same person under a different romanization, more narrowly: the company systematized complex agent tasks as Pro and large-scale automation as Mini, and will let customers choose based on work and throughput. The DigitalToday quote does not include "every competing model." Do not merge the two sentences into a stronger claim than either page makes.

Neither article prints the index scale, the date of the Artificial Analysis run, or the prompt settings. A score of 24 is not interpretable as a percent. It is not comparable, on these pages, to coding benchmarks other labs publish under different names. The on-premise pitch also depends on quantization. Both pages say a single GPU is enough when quantization is applied. They do not say which GPU, which bit width, or what quality drop that quantization causes.

The 70 percent API discount is only in Chosun Biz, which says the cut runs "through the 10th" and does not print a clock time. DigitalToday does not mention a discount at all. If you are checking a bill, that promotion is a single-source line.

Where the second page only repeats

On the core product facts, DigitalToday repeats Chosun Biz: 35 billion parameters, 3 billion active, a mixture-of-experts layout, an Intelligence Index score of 24, single-GPU on-premise use, a Pro versus Mini split, and an OpenRouter API. It does not add a second measurement. That is why the confirmation count on this piece stays modest. The second page is useful as a check that those integers were not a one-off typo. It is not an independent benchmark lab.

What DigitalToday does not carry is the list of named rivals and the promotion. What Chosun Biz does not carry is the stray "Wednesday." Treat the shared integers as the sturdy part, and treat the superlatives as Upstage's wording reported by one desk.

What to do as a reader (not a trade)

This is not investment advice. Solar Mini 4 is a tool choice, not a signal to trade a supplier's shares. If your stack already uses Solar Pro 4 on OpenRouter, the practical question is whether a cheaper, smaller active model can take the repetitive calls. The pages do not publish a price card beyond the 70 percent promotion on one of them, so "cheaper" is not yet a number you can put in a budget from these stories alone.

Ask Upstage, or read the console docs, for four items the news pages skip: the context length, the rate limits, which languages the model is reliable in, and what the score of 24 was measured on. Until those exist, the class comparison with named rivals is a press claim. Run one repetitive workflow you already have, with quantization if you plan to self-host, and compare failure rate against the Pro model on the same inputs. A single GPU is only a win if the answers stay acceptable.

For other model launches in the same few days, Gemini 4 Argon's rollout and Cohere's Embed 5 pricing story are different products with different jobs. They do not validate an Intelligence Index score of 24. Chosun Biz's October 1 story is the fuller of the two pages. DigitalToday's shorter item repeats the headline specifications and leaves the rival list and the discount out.