韩国 Upstage 出的 Solar Mini 4,3B 激活参数跑出 208 tokens/秒,在 Cline 里免费能用,写代码的可以试试。
#Solar Mini 4
Upstage 的 Solar Mini 4 这两周在 Nous Portal 免费用,3B 激活参数却能打过 10 倍大的模型,赶紧去试试。
Upstage的Solar Mini 4虽参数量大但激活参数少,性能介于Qwen 3.6和3.8之间,专为自动化任务设计。
Solar Mini 4 establishes a new Pareto optimal point for models under 3B active parameters, offering strong long-context reasoning at 83% on AA-LCR v1.1. Developers should note its agentic coding weakness at 1% on Terminal-Bench 4.0 and consider its 5x higher task cost compared to GPT-6 Luna (max) despite similar per-token pricing.
This model provides a specialized alternative to general-purpose LLMs for structured decision-making tasks. Developers working on applications requiring consistent decision outputs can now access a cost-effective solution priced at $0.10/M input tokens. Compared to other decision-focused models, Solar Mini 4 offers a more affordable entry point for businesses implementing AI-driven decision systems.
The model's mixture-of-experts architecture delivers high performance with lower costs, making it accessible for developers building agentic applications. For developers in Southeast Asia, this provides a cost-effective alternative to larger models while maintaining strong capabilities for tool integration and reasoning tasks.
UpstageAI刚在OpenRouter上线两款模型,限时半价,值得关注。