Upstage / upstage/solar-mini4

Solar Mini 4 - access through LLMTR

Solar Mini 4 is Upstage's compact model for agentic work where response speed and cost decide: 35B total parameters with 3B active per step. For high-volume, cost-sensitive agent work it costs less than Solar Pro 4. It has a 524,288-token context window; Upstage states an output size of up to 128K tokens per response, and generated tokens come out of that same window. Reasoning is off by default. You select a level with `reasoning_effort` across none, minimal, low, medium, high, xhigh, and max; the model starts thinking at `low`, while `none` and `minimal` are accepted but produce no reasoning tokens at all. When reasoning is on, the trace is returned separately in the response's `reasoning` field and streamed incrementally. Every `tool_choice` value works for native tool/function calling, and it can emit several tool calls in one response. It produces JSON object output and structured output that conforms exactly to a supplied schema. Repeated prompt prefixes engage the prompt-cache discount. It accepts text only; image and audio input are not supported. Official language coverage is Korean, English, and Japanese. Reasoning tokens count as output and are drawn from the `max_tokens` budget, so keep it generous while reasoning is on.