Sign in
Pending benchmarksTracking since Nov 1, 2026

Is Solar Mini 4 state of the art?Pending benchmarks.Announced Sep 22, 2026; no independent results yet.

Upstage's compact model for high-volume agent work: a 35B mixture-of-experts model pretrained from scratch by Upstage, with 3B parameters active per token. It has a 512K-token context, up to 128K output tokens, optional reasoning, structured outputs and tool calling, in Korean, English and Japanese. Available through the Upstage Console API, Solar Chat and OpenRouter, with on-premises deployment. No weights published.

Announced facts

The details so far

Lab
Upstage
Announced
Sep 22, 2026
API
Live · solar-mini4 · Upstage Console, model solar-mini4 (snapshot solar-mini4-260922), OpenAI-compatible at api.upstage.ai/v1; also on OpenRouter as upstage/solar-mini4
Open weights
Closed · Not published; no Solar Mini 4 repo on Upstage's Hugging Face org as of Nov 1, 2026. On-premises deployment is arranged through Upstage
Parameters
35B total, 3B active per token (mixture of experts)
Type
Text model for agent workloads: retrieval, classification, structured output and tool use; reasoning off by default, set with reasoning_effort
Architecture
Mixture of experts, 35B total and 3B active per token, pretrained from the ground up by Upstage
Context
512K tokens (524,288), up to 128K output tokens
Languages
Korean, English and Japanese
Training data cut-off
February 2026
Capabilities
Chat, reasoning, JSON mode, JSON Schema structured outputs, tool calling and parallel tool calling
Price
$0.10 per 1M input tokens, $0.01 per 1M cached input tokens and $0.40 per 1M output tokens (Upstage standard pricing); launch discounts ran through Oct 22, 2026 (UTC)
Availability
Upstage Console API and Playground, Solar Chat, OpenRouter and on-premises deployment; free on Nous Portal (Hermes Agent) for two weeks from Oct 5, 2026, per Upstage
Related
Upstage's solar-decide decision model is built on Solar Mini 4 and is a separate product

Tags marked unsure are not stated on the lab's own pages.

Next steps

What would let us call it

Verdicts use only independent boards we already track. Two consecutive snapshots are required before any crown.

  • Publicly announced with a source
  • Usable by the public (API or weights)
  • Listed on an independent board we track
  • Enough results for a rank in a category
  • Holds across two snapshots (crowns only)
Self-reported

What the lab claims

Self-reported. Shown for context; not used in any verdict.

Lab claim: 24.1 on the Artificial Analysis Intelligence Index v4.3.2, the highest among the 3B-active models in Upstage's comparison, ahead of Qwen3.6 35B A3B (18) and Nemotron 3.5 Lightning (13).

Lab claim: 22.3% on AutomationBench-AA, 47.2 on τ³-Banking, 83.3% on AA-LCR, 47.6% on SciCode and 25.8% on Humanity's Last Exam.

Lab claim: over 70 tokens per second per request at 32 concurrent requests on two H100 GPUs, and the quantized model runs on a single H100 80GB.

Lab claim: in a small internal test of three Korean-language agent tasks, about 43% lower cost per completed task than MiMo-V2.5 ($1.25 vs $2.20 per 1,000 task sets).

Receipts

Sources

LABSolar Mini 4 model docs (Upstage Console)console.upstage.ai/docs/models/solar-mini-4LABSolar Mini 4: Built for High-Volume Agent Workloads (Upstage blog)upstage.ai/blog/en/solar-mini-4 · Oct 1, 2026LABUpstage API pricingupstage.ai/pricingLABUpstage post on X (solar-decide on Solar Mini 4)x.com/upstageai/status/2104717383910977807 · Sep 28, 2026LABUpstage announcement on Xx.com/upstageai/status/2105934955314663673 · Oct 2, 2026LABUpstage post on X (free on Nous Portal)x.com/upstageai/status/2107138817161646460 · Oct 5, 2026

Tracking continues until an independent board lists it.