Upstage unveiled Solar Mini 4 on Wednesday, pitching the 35-billion-parameter model to Korean enterprises that want on-premise automation without reserving a multi-GPU cluster for every workflow.
Architecture and benchmarks
The model uses a mixture-of-experts design that activates about three billion parameters per inference step, cutting compute while keeping a larger knowledge pool. Upstage said it scored 24 points on Artificial Analysis' Intelligence Index, a vendor-neutral scoreboard it cites to argue the Mini line punches above other small models.
Quantization allows a single graphics card to host the model, a constraint many mid-sized manufacturers and public agencies cite when they reject cloud-only AI contracts.
Go-to-market split
CEO Kim Sung-hoon said Upstage will position Solar Pro for agentic tasks that need complex judgment and Solar Mini for repetitive, high-volume jobs such as document classification or invoice extraction. Solar Mini 4 follows August's Solar Pro 4 listing on OpenRouter and competes with open weights from Google Gemma and Alibaba Qwen that dominate budget-conscious pilots.
Upstage is also building English exam items for Korea's college entrance test and running a cybersecurity consortium with SK Telecom, but Solar Mini 4 is the revenue workhorse aimed at companies that told the startup they "do not have enough GPUs" for full-size models.
Deployment scenarios
Manufacturers testing Solar Mini 4 typically start with invoice OCR and purchase-order matching, workloads that need high throughput but tolerate occasional errors. Upstage said quantization presets will ship for common NVIDIA data-center cards used in Korean private clouds.
Public-sector pilots through the Wavus GeoAI partnership could follow, pairing spatial data with Mini models that run inside agency firewalls. That aligns with Upstage's separate MOU to build English exam items for the College Scholastic Ability Test — both projects reward models that are cheap to host.
Risks and limits
Mixture-of-experts models can be harder to debug when experts specialize unexpectedly. Upstage will need transparent logging for regulated industries. Competitors argue that open weights from U.S. and Chinese labs reduce switching costs, pressuring Upstage to bundle services and support rather than weights alone.
Still, for CIOs who face GPU rationing, a 35-billion-parameter model that fits one card is an easier budget conversation than a frontier cluster.