Upstage Unveils 'Solar Mini 4' with Active 3B MoE Architecture

Upstage has released Solar Mini 4, a model independently pre-trained and optimized for agentic tasks. It employs an MoE architecture that activates only 3B parameters per token out of a total of 35B, delivering high efficiency. In the Artificial Analysis Intelligence Index v4.3.2 evaluation, it recorded the highest overall performance among active 3B models with a score of 24.1.

Performance and Benchmarks

Compared to Gemma 4 31B, which has 31B active parameters, Solar Mini 4 activates only about one-tenth of the parameters yet scores 9.1 points higher on the AA Index. It also recorded an overall score 1.1 points higher than Nemotron 3 Ultra, which has 55B active parameters. In detailed evaluations, it scored 83.3 on long-context reasoning (AA-LCR) and 47.6 on code-based scientific problems (SciCode).

It recorded 22.3% on AutomationBench-AA, which evaluates the ability to perform tasks using actual SaaS tools, and 47.2 on the τ³-Banking evaluation. It supports tool calling and parallel tool calling, enabling the handling of complex workflows.

Cost Efficiency

Designed to lower the cost of repetitive agentic tasks. In internal tests, when performing the same agentic tasks with Solar Mini 4 and MiMo-V2.5, the cost to complete 1,000 sets was $1.25 for Solar Mini 4 and $2.20 for MiMo-V2.5, approximately 43% lower. Both models passed quality checkpoints.

Even when handling 32 concurrent requests on two H100 GPUs, it maintained a generation speed of over 70 tokens/s per request. For on-premises deployment, quantized models can run starting from a single H100 80GB GPU.

Pricing and Availability

Available via Upstage Console API, Solar Chat, and OpenRouter, with support for on-premises deployment. Standard API pricing is $0.10 per million input tokens, $0.01 for cached input, and $0.40 for output. Upstage Console offers a 70% discount until October 10 UTC.