Solar Mini 4 is Upstage's compact language model and the small sibling of Solar Pro 4. It is a mixture-of-experts model with 35 billion parameters in total but only 3 billion active on any given token, and that ratio is the whole design: a request is priced and served roughly like a 3B model while the model can still draw on the knowledge held across all 35B. Upstage pairs it with a 524,288-token context window — the same window the larger Solar models get — and aims it squarely at agentic use.
On AI/ML API it is served as upstage/solar-mini4 through the standard chat completions endpoint, so any OpenAI-compatible client works with only the base URL changed.
Prices per million tokens are hard to feel, so here is the same number in work units. A support-style exchange of roughly 1,000 input and 300 output tokens costs about $0.0003 — around 3,400 such calls per dollar. A retrieval-heavy request that stuffs 50,000 tokens of context in and writes 1,000 back costs about $0.0074, so a dollar buys roughly 135 of them.
The cached-input rate is the one worth designing around: at $0.013754 it is a tenth of the standard input price. An agent that replays the same system prompt and tool definitions on every step — which is what agents do — pays a tenth for that repeated prefix once it is cached. On a loop with a 20,000-token fixed preamble and thirty steps, that is the difference between about $0.083 and $0.008 per task in prefix cost alone.
A dense 35B model would have to move all 35 billion parameters for every token it generates. Solar Mini 4 routes each token to a small subset of experts, so only about 3 billion take part. The practical consequences are the ones that show up in a bill and in a latency graph: cost per token lands in small-model territory, and throughput stays high enough for loops that make many short calls rather than a few long ones.
What the ratio does not buy is frontier-level depth on hard single-shot reasoning. A 3B active-parameter forward pass is a 3B forward pass, whatever the total parameter count — which is why Upstage positions this model for volume and agentic work rather than as a replacement for its flagship tier.
524,288 tokens is roughly 1,500 pages of plain prose, or a mid-sized codebase with its tests and documentation. Two things follow. First, a whole class of retrieval plumbing becomes optional: for corpora that fit, the model can be handed the documents directly instead of a chunking pipeline, an embedding store and a reranker. Second, an agent's transcript does not have to be summarised and truncated every few steps, which is where long agent runs usually start losing the thread.
The output ceiling is 131,072 tokens, so the model can also write at length — a full refactor across many files, or a long structured report, without splitting the job into several calls and stitching the results.
High-volume agentic work. Pipelines that run dozens of steps per task, where the per-step price decides whether the workflow is viable at all. Parallel tool calls let one step issue several tool invocations at once rather than serialising them.
Document processing at scale. Summarising, extracting and classifying across long documents, where the context window removes the chunking layer and the price makes running over a whole archive realistic.
Routine engineering tasks. Code review comments, migration edits, test generation and other work that is well specified and repetitive — the kind of job usually handed to a frontier model out of habit rather than need.
Structured extraction. Turning free text into JSON against a schema, at a price where running it over millions of records is not a budget conversation.
Solar Mini 4 is text-only: no image or audio input. If a request involves screenshots, diagrams or scanned pages, Solar Pro 4 handles vision and adds web search. For problems where a single hard answer matters more than the cost of getting it — novel algorithmic work, intricate multi-step proofs, deep legal or medical analysis — a frontier-tier model is the better trade even at ten times the price. And for very short, latency-critical classification, a model with a smaller context window may return faster simply because there is less machinery involved.
Solar Pro 4 is the larger sibling: same 524,288-token window and 131,072-token output ceiling, but with image input, vision and web search on top, at $0.041262 / $0.165048 per 1M tokens. Solar Pro 3 is the previous generation, worth comparing when an existing integration is already pinned to it.
The honest summary: pick Solar Mini 4 when the workload is text-only and the number of calls is large; pick Solar Pro 4 when any request might carry an image or need the open web. Both share the same context window, so the choice is about modality and depth rather than about how much you can fit in a prompt.
Measured on the AI/ML API gateway across live traffic. Figures are medians unless stated. Last updated 2026-10-12.
Agentic pipelines. Multi-step loops that plan, call tools and check their own work. Parallel tool calls mean a single step can fire several tool invocations at once, and the cached-input rate keeps the repeated system prompt from dominating the bill across thirty or forty steps.
Long-document analysis. Contracts, filings, research papers and support archives go into a single request rather than through a chunk-embed-rerank pipeline, because 524,288 tokens covers roughly 1,500 pages.
Repository-scale code work. Code review, test generation and mechanical migrations across a mid-sized codebase, with room for a 131,072-token reply when the answer is a large diff rather than a comment.
Structured extraction at volume. Turning invoices, tickets or product feeds into schema-conformant JSON, at a price where running the job across millions of records is affordable.
import requests r = requests.post( "https://api.aimlapi.com/v1/chat/completions", headers={"Authorization": "Bearer " + AIMLAPI_KEY}, json={ "model": "upstage/solar-mini4", "messages": [ { "role": "user", "content": "Hello!" } ] }, ) print(r.json())
const r = await fetch("https://api.aimlapi.com/v1/chat/completions", { method: "POST", headers: { Authorization: `Bearer ${process.env.AIMLAPI_KEY}`, "Content-Type": "application/json", }, body: JSON.stringify({ "model": "upstage/solar-mini4", "messages": [ { "role": "user", "content": "Hello!" } ] }), }); console.log(await r.json());
curl -X POST https://api.aimlapi.com/v1/chat/completions \ -H "Authorization: Bearer $AIMLAPI_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"upstage/solar-mini4","messages":[{"role":"user","content":"Hello!"}]}'
OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Input | |
| Output | |
| Cached input |
Agentic pipelines. Multi-step loops that plan, call tools and check their own work. Parallel tool calls mean a single step can fire several tool invocations at once, and the cached-input rate keeps the repeated system prompt from dominating the bill across thirty or forty steps.
Long-document analysis. Contracts, filings, research papers and support archives go into a single request rather than through a chunk-embed-rerank pipeline, because 524,288 tokens covers roughly 1,500 pages.
Repository-scale code work. Code review, test generation and mechanical migrations across a mid-sized codebase, with room for a 131,072-token reply when the answer is a large diff rather than a comment.
Structured extraction at volume. Turning invoices, tickets or product feeds into schema-conformant JSON, at a price where running the job across millions of records is affordable.
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
Solar Mini 4 This page | Agentic workflows on a budget | |||
| Agentic workflows, coding & document analysis | ||||
| Reasoning + agents | ||||
| Fast, high-volume tasks |
Solar Mini 4 is Upstage's compact language model: a 35-billion-parameter mixture-of-experts with 3 billion active parameters and a 524,288-token context window. It is built for agentic use cases where cost per step matters.
On AI/ML API, Solar Mini 4 costs $0.13754 per 1M input tokens and $0.55016 per 1M output tokens. Cached input is billed at $0.013754 per 1M tokens.
Solar Mini 4 accepts up to 524,288 tokens in a single request and can return up to 131,072 tokens, so long documents and large codebases fit without chunking.
It is a text-in, text-out model with reasoning, streaming, structured outputs, tool calling and parallel tool calls — the pieces an agent loop needs to plan, call tools and verify its own work.
Send a POST request to https://api.aimlapi.com/v1/chat/completions with upstage/solar-mini4 as the model id and your AI/ML API key in the Authorization header. The endpoint is OpenAI-compatible, so existing chat completion clients work unchanged.
