Upstage's compact proprietary model. Its API documentation calls it a "cost-efficient, compact language model (35B total / 3B active parameters), built for agentic use cases where response speed and cost matter." Apart from size, its published specifications match Solar Pro 4's: a 512K-token context window (524,288 tokens on OpenRouter), up to 128K output tokens, a reasoning mode, structured outputs and tool calling, English, Korean and Japanese, and a February 2026 training-data cutoff. It became generally available on September 22, 2026 (model ID solar-mini4-260922, alias solar-mini4) on the Upstage Console and Playground and for on-premises deployment. CEO Sung Kim had opened a free preview on September 13. The console lists it at $0.10 per million input tokens and $0.40 per million output tokens, with 50% off through October 22. No weights are released. It replaces the 10.7B-parameter Solar Mini API model (solar-mini-250422), which Upstage retires on October 30, 2026 (KST), telling users to migrate to the solar-mini4 alias.

Upstage has published no technical report, benchmark table or training details. OpenRouter's listing calls the model a mixture-of-experts. AI Times, relaying Kim's preview announcement, reported that it carries over the architecture and training data validated in Solar Open 2 and performs close to Solar Pro 4. Upstage has not said whether the model was pretrained from scratch. The 35B total / 3B active shape is also exactly that of the open Qwen3.5/3.6-35B-A3B checkpoints, and earlier Solar models were built on external bases. Given both, the 35B/3B figures stay out of the structured scale fields and the model is marked as not established from scratch. The only outside numbers so far come from a Sogang University team that tested the preview on 46 Korean CSAT math questions with Python tool use, as reported by AI Times. It solved 93.5% of the questions within three attempts and 76.8% on the first attempt, against 87.7% for Solar Pro 4. It averaged 14.9 seconds per attempt, against 120.5 seconds for Solar Pro 4. The team cautioned that the models were reached through different access routes and infrastructure.

Artificial Analysis lists it as a reasoning model (release date September 22) and scores it 24 on Intelligence Index v4.3.2. On the same index version that places it just below Solar Open 2 and Solar Pro 4, roughly consistent with Upstage's "close to Solar Pro 4" pitch for a model at a fraction of the price. AA records a 1,048,576-token context; the figure here stays at Upstage's documented 512K.

Upstage launched Solar Decide (beta) the same day, a "structured decision model, served as a System One endpoint on Solar Mini 4" (POST /v1/systemone). Given a state and typed questions, it returns a choice, a score or a yes/no answer with a calibrated probability. It uses one forward pass per decision instead of generating text.

Model Details

Architecture MOE
Context window 524,288
AA Intelligence 24 on v4.3.2
License Proprietary
enterpriseagenticreasoningmultilingualefficiencymoe

Related