Upstage: Solar Mini 4

Solar Mini 4 is a text language model from Upstage, listed on api14 under the model id upstage/solar-mini4. Upstage describes it as a compact, cost-efficient model built as a 35B-parameter mixture-of-experts with 3B active parameters. Its provider release date is 2026-09-23. The design goal stated by the provider is agentic use cases, where the model is expected to act on instructions and call tools rather than only produce prose.

The model accepts text input and returns text output, so it is a pure text-to-text system with no image or audio handling. Its context window is 524,288 tokens, which is large enough to hold long documents, extended conversation histories or sizeable tool outputs in a single request. The maximum output length is 131,072 tokens. Because it is a mixture-of-experts model with 3B active parameters, only a fraction of the 35B total parameters is engaged per token, which is the mechanism behind the provider's cost-efficiency claim.

Those capabilities point to a few practical patterns. The long context suits retrieval-style workloads where you paste in many source documents and ask for a synthesis, and it also suits agent loops that accumulate tool results over many turns without needing to trim history. Tool calling and structured outputs make it a reasonable fit for pipelines that need machine-readable responses, such as extracting fields into a schema or routing a request to a downstream function. Reasoning controls let you trade latency and verbosity against answer quality on harder tasks, while the compact active parameter count keeps it viable for high-volume, lower-complexity jobs.

You call it through api14's OpenAI-compatible API using the model id upstage/solar-mini4. The supported request parameters are frequency_penalty, include_reasoning, max_tokens, parallel_tool_calls, presence_penalty, reasoning, reasoning_effort, response_format, structured_outputs, temperature, tool_choice, tools and top_p. In practice that means you can set sampling behaviour with temperature and top_p, discourage repetition with the two penalty parameters, cap the reply with max_tokens, request JSON or a schema via response_format and structured_outputs, and wire up function calling with tools, tool_choice and parallel_tool_calls. The reasoning, reasoning_effort and include_reasoning parameters control whether and how much intermediate reasoning is produced or returned.

The facts given do not include benchmark results, latency figures or quality comparisons, so there is no basis here for claiming it outperforms or underperforms other models. The clearest boundary is modality: it handles text only, so a task that requires reading images or audio needs a different model. If you need the largest possible output for a single very long generation, note that output is capped at 131,072 tokens even though the context window is 524,288 tokens. Within text work, the combination of a large context, tool calling and structured outputs is what makes it worth trying first for agentic and extraction pipelines.