Upstage launches Solar Mini 4, a compact MoE model with 3 billion active parameters

Upstage has released Solar Mini 4, a 35-billion-parameter Mixture-of-Experts model that activates only 3 billion parameters per forward pass and offers a 524-thousand-token context window. The model targets agentic workloads where response speed and cost are the primary constraints, and it supports tool calling, structured outputs and optional reasoning.

Architecture and performance

The MoE design keeps the theoretical model size large while consuming resources comparable to a much smaller model. According to company data, throughput reaches 83 tokens per second (P50, best among providers tested) with latency of 0.36 seconds (P50, best provider). Those figures place the model at the fast end of its category, which matters especially when it serves as the planning and control layer in long agentic loops.

Pricing and context window

Pricing is set at $0.05 per million input tokens and $0.20 per million output tokens (approximately 0.18 and 0.73 shekels respectively). Combined with the 524-thousand-token context window, the model allows long conversation histories, multiple documents or complex environment state to be streamed without hitting a length limit — a significant advantage in agentic workflows where context accumulates over time.

Multilingual support and agentic capabilities

The model offers fluent Korean alongside strong English and Japanese, reflecting Upstage's regional focus. Beyond language, built-in support for tool calling and structured outputs simplifies integration with external systems, databases, APIs and execution environments without additional parsing and error-correction layers. The optional reasoning toggle adds flexibility for tasks that require multi-step planning.

Availability, ZDR and load balancing

The model is available through the Upstage API console starting 23 September 2026. Using a BYOK (Bring Your Own Key) setup with enforced ZDR (Zero Data Retention) requires the organisation in Upstage to be ZDR-enabled; customers must contact the company to enable it. The system includes automatic routing to a healthy provider when a top-tier provider errors, provided the request filters allow it; per-provider availability data is accessible via API endpoints.