Upstage: Solar Mini 4

The myaiexp write-up covers Upstage's Solar Mini 4 released 2026-09-22 as solar-mini4-260922 (alias solar-mini4), framing it as the compact tier of the Solar family with a 35B mixture-of-experts architecture, 3B active parameters, a 512K context window, up to 128K max output, and a training data cutoff of February 2026 Pricing is documented with a launch promotion running until 2026-10-23 UTC: input drops from $0.10 to $0.05 per million tokens, cached input from $0.01 to $0.005, and output from $0.40 to $0.20 per million tokens. The piece explicitly positions Solar Mini 4 for high-frequency agent work such as routing, document classi

Benchable documents Upstage's Solar Mini 4 (released 2026-09-23) as a 35B-parameter mixture-of-experts model with 3B active parameters and a 524K context window, with pricing at $0.05 per million input tokens, $0.20 per million output tokens, and $0.005 per million cached-input read tokens. Supported parameters include Performance assessment places Solar Mini 4 in the 96th percentile for speed across eight benchmarks and the 88th percentile for price competitiveness, with a 99% reliability success rate. Accuracy is uneven: strong in Ethics (99%) and Email Classification (96%), solid General Knowledge at 93%, but weaker in Hallucinati

Upstage's Solar Mini 4 is a compact, cost-efficient language model positioned as a 35B-parameter mixture-of-experts with 3B active parameters and a 524K context window, released on 2026-09-23 and ranked 238 of 435 in coding on the LMC scorecard. The aggregator reports an Overall Score of 40/100 with strong sub-signals API pricing is listed at $0.05 per million input tokens and $0.20 per million output tokens, with a maximum output of 131.1K tokens from the 524.3K context window. The page frames the model as built for agentic use cases where response speed and cost per call dominate, though accuracy in coding, reasoning, and mathemat