
Solar Mini 4: Specs, Pricing, Benchmarks & Agent Fit
By Julian Brooks
MyClaw Editorial
MyClaw
Run Best-in-Class AI Agents Now
Run OpenClaw or Hermes Agent on managed MyClaw hosting, so your AI agent stays online, updated, and ready for real work.
AI Takeaway
- What is Solar Mini 4? Upstage’s compact proprietary text model uses a 35B-parameter mixture-of-experts design with 3B active parameters, a 512K context window, and up to 128K output tokens.
- How much does it cost? List pricing is $0.10 per 1M input tokens, $0.01 cached input, and $0.40 output, excluding VAT. A 50% launch offer runs through October 22, 2026 UTC.
- What is it best for? High-volume text workflows that need tool calling, structured output, Korean/English/Japanese support, or long inputs—but not native image, audio, or video work.
- Is its performance proven? Not by a first-party benchmark suite. Early third-party tests are fragmented, so test your own tasks before replacing a larger model.
- Can it power an OpenClaw agent? Yes, through Upstage or a compatible provider such as OpenRouter. Solar Mini 4 is the intelligence layer; the agent runtime supplies tools, memory, and ongoing execution.
Solar Mini 4 Specs at a Glance
Upstage launched Solar Mini 4 in September 2026 as the compact member of its current Solar family. It pairs a large context window and agent-oriented API features with sparse activation.
| Specification | Solar Mini 4 |
|---|---|
| Architecture | Mixture of experts |
| Parameters | 35B total, 3B active |
| Context window | 512K tokens (524,288 exact) |
| Maximum output | 128K tokens (131,072 exact) |
| Training cutoff | February 2026 |
| Input and output | Text |
| Languages | Korean, English, Japanese |
| Agent features | Tool calling, structured output, optional reasoning |
| Upstage IDs | solar-mini4, solar-mini4-260922 |
| OpenRouter ID | upstage/solar-mini4 |
| Access | Upstage API, OpenRouter, and on-premises deployment |
| Weights | Proprietary; no open release announced |
That distinction between model and runtime matters. An LLM generates the next response; an agent surrounds it with goals, tools, memory, and a loop. This AI agent vs. LLM guide explains why an inexpensive model can still need substantial infrastructure around it.
What 35B Total and 3B Active Means
A mixture-of-experts model contains specialized parameter groups but routes each token through only part of the network. Solar Mini 4 has 35 billion parameters in total while activating about 3 billion per token. That can reduce inference work and helps explain the low API price, but it does not prove the reasoning quality of a dense 35B or larger frontier model. Parameter counts describe architecture, not task success.
Why 512K and 524,288 Tokens Are Both Correct
Upstage rounds the context window to 512K and the maximum output to 128K. The exact limits are 524,288 context tokens and 131,072 output tokens because each “K” represents 1,024 tokens. These figures describe the same limits, not two editions.
A large window also does not guarantee perfect recall across a long document. Test whether the model can find evidence near the beginning, middle, and end of your actual inputs. Long-context reliability matters more than the maximum number printed on the model card.
Solar Mini 4 Pricing and API Access
Solar Mini 4 has a low list price, but its temporary launch discount can make comparisons confusing.
| Token type | Standard price per 1M | Launch price through October 22, 2026 UTC |
|---|---|---|
| Input | $0.10 | $0.05 |
| Cached input | $0.01 | $0.005 |
| Output | $0.40 | $0.20 |
Upstage states that these prices exclude 10% VAT. Date-stamp any budget built around the promotional rate; a workload that looks unusually cheap in October may cost twice as much after the offer ends.
At standard rates, a run with 100,000 input tokens and 5,000 output tokens costs about $0.012 before caching, reasoning, retries, and tax. The launch rate cuts that to about $0.006. In production, failed tool calls and repeated attempts can matter more than the nominal token price.
Upstage and OpenRouter Model IDs
Use solar-mini4 when you want Upstage’s stable alias to follow the current supported version. Use solar-mini4-260922 when you need to pin the initial snapshot for reproducible testing. OpenRouter exposes the model as upstage/solar-mini4.
Both routes expose OpenAI-compatible chat-completion patterns, but the provider remains part of the decision. Check rate limits, retention controls, regional availability, and billing terms for the route you choose. Reasoning is off when omitted. Upstage accepts low, medium, high, xhigh, or max to enable it; none and minimal keep it off.
Is Solar Mini 4 Good for AI Agents?
Solar Mini 4 is designed first for frequent, constrained calls, not as a proven replacement for premium models in long autonomous runs. Candidate tasks include classification, extraction, routing, summarization, schema-constrained responses, and simple tool selection. Korean-first workflows are especially relevant, while English and Japanese support suits mixed-language operations.
The model is text-only. If an agent must inspect screenshots, listen to calls, or reason directly over video, it needs a separate multimodal model or a preprocessing tool that converts the source into text.
For OpenClaw, the right model is determined by the workflow rather than a leaderboard headline. Tool selection, valid arguments, instruction retention, and recovery from errors all affect whether a run finishes. A broader OpenClaw model selection guide can help define those criteria before testing Solar Mini 4.
Solar Mini 4 Benchmarks: What Is Actually Published?
Upstage’s launch materials do not include a dedicated Solar Mini 4 benchmark table. Early third-party tests use different prompts and methods, while major leaderboards do not yet offer a consistent quality score. Treat those results as signals, not proof that the model will complete your workflow.
That leaves several important questions for workload testing:
- Does it select the correct tool and supply valid arguments?
- Does structured output remain valid across varied inputs?
- Can it follow instructions deep inside long context?
- How often does a failed step require a retry or stronger fallback?
- What are the latency and total cost per completed task?
Measure successful outcomes, not isolated calls. A model that costs one-third as much but needs four attempts is not the cheaper production choice.

Solar Mini 4 vs Solar Pro 4
Solar Mini 4 and Solar Pro 4 expose similar limits: both support 512K context, up to 128K output, tool calling, structured output, and optional reasoning. Their vendor positioning differs: Mini 4 targets high-frequency, cost-sensitive work; Pro 4 targets harder multi-step agent tasks.
| Decision point | Solar Mini 4 | Solar Pro 4 |
|---|---|---|
| List input / output price | $0.10 / $0.40 per 1M | $0.30 / $1.20 per 1M |
| Context / max output | 512K / 128K | 512K / 128K |
| Best starting workload | Repetitive, structured, high-volume | Complex, multi-step, costly-to-fail |
| Published model-specific evidence | Limited at launch | Broader vendor evaluation claims |
| Practical role | Economical default candidate | Escalation or difficult-task candidate |
Solar Pro 4’s list token rates are three times Mini 4’s, but price per token is not price per result. Run both on the same prompts, tools, and success criteria. Keep Mini 4 on routine tasks it completes cleanly; escalate work that requires planning, recovery, or higher confidence.
Model choice is also separate from who operates the runtime. This MyClaw vs. OpenClaw comparison covers managed hosting versus self-hosting, while a disciplined OpenClaw multi-agent setup can apply different model policies to separate roles.
How to Test Solar Mini 4 in a MyClaw OpenClaw Workspace

MyClaw provides managed hosting for private, always-on OpenClaw agents. It does not replace Solar Mini 4; it gives the model a persistent environment for tools, memory, files, and scheduled work.
- Connect the provider. Start a MyClaw OpenClaw workspace and add an OpenRouter key or another compatible endpoint that exposes Solar Mini 4.
- Run one fixed workflow. Select
upstage/solar-mini4, keep the prompt, tools, inputs, and success criteria unchanged, and test a repeatable low-risk job. - Route by results. Compare completion rate, tool accuracy, latency, retries, and total token cost; keep Mini 4 for economical passes and send difficult failures to a stronger model.
MyClaw hosting and third-party model usage are separate costs, so check the current MyClaw hosting plans alongside the API bill.
Verdict: Who Should Choose Solar Mini 4?
Choose Solar Mini 4 when you need a low-cost text model for frequent structured tasks, very long inputs, or Korean-centered multilingual workflows. Pair it with a stronger fallback when errors are costly, and skip it when the task requires native image, audio, or video understanding.
Its price and specifications earn it a serious test, not an automatic production promotion. Use a representative task set, track cost per successful completion, and change routing only when the results hold up under real tools and real failure cases.
Skip the Setup, Run Best-in-Class AI Agents Now
Launch a managed OpenClaw or Hermes Agent workspace in minutes, with always-on hosting, updates, and support handled by MyClaw.