Upstage Releases Solar Mini 4: 35B MoE Model with 524K Context at $0.05/$0.20 per Million Tokens
Upstage has released Solar Mini 4, a compact mixture-of-experts model with 35B total parameters, 3B active parameters, and a 524K token context window. The model targets agentic workloads and is priced at $0.05 per 1M input tokens and $0.20 per 1M output tokens, a promotional 50% discount off standard rates.
Solar Mini 4 at a glance
How Solar Mini 4 ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| Speed (tok/s) | 131 tok/s | 48 of 127 | Artificial Analysis, Oct 2026 ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
Upstage Releases Solar Mini 4
Upstage has released Solar Mini 4, a compact mixture-of-experts (MoE) language model built for cost-sensitive, agentic workloads. The model is now available through OpenRouter and Upstage's API.
Specifications
Solar Mini 4 uses a 35-billion-parameter MoE architecture with only 3 billion parameters active per forward pass, a design intended to keep inference costs and latency low while retaining broad model capacity. It ships with a 524,288-token (524K) context window, positioning it for long-document and long-horizon agentic tasks.
According to Upstage, the model offers fluent Korean alongside strong English and Japanese performance, continuing the company's focus on multilingual support centered on Korean-language use cases.
Pricing
Solar Mini 4 is listed at $0.05 per 1 million input tokens and $0.20 per 1 million output tokens on OpenRouter — a promotional 50% discount off the standard listed rate of $0.10 / $0.40 per 1 million tokens. Cache-read pricing is listed at $0.01 per 1M tokens (input) and $0.005 per 1M tokens (output) under the discounted tier.
OpenRouter data shows throughput up to 95 tokens per second and P50 latency as low as 0.28 seconds across providers, with 100% uptime recorded over the most recent 24-hour and 3-day windows.
Deployment and Access
Upstage supports bring-your-own-key (BYOK) access with zero-data-retention (ZDR) enforcement. Organizations wanting ZDR enabled on their Upstage Console API key must contact Upstage directly to activate it; without this, standard data-retention terms apply.
Model Family Context
Solar Mini 4 joins two other current Upstage releases available on OpenRouter: Solar Pro 4, which shares the same 524K context window and is priced at $0.09 / $0.36 per 1M tokens for office- and coding-oriented workflows, and Solar Pro 3, a larger 102B-parameter MoE model (12B active) with a 131K context window priced at $0.15 / $0.60 per 1M tokens. Solar Mini 4 sits below both in parameter count and price, targeting lighter-weight, higher-volume use cases.
No benchmark scores (e.g., MMLU, HumanEval) or a training data cutoff date have been disclosed by Upstage for this release.
What This Means
Solar Mini 4's combination of a 524K context window and 3B active parameters puts it in a niche aimed squarely at high-throughput agentic pipelines — think document retrieval, multi-step tool use, and long-running conversational agents — where per-token cost and latency matter more than raw benchmark ceiling. The MoE design (35B total, 3B active) is a common strategy for keeping inference costs down while preserving some of the capacity benefits of a larger dense model, though without published benchmark scores it's not possible to independently verify how it compares to similarly priced competitors like smaller Gemini, Llama, or Qwen variants. Upstage's continued emphasis on Korean-language fluency, paired with English and Japanese support, signals the company is prioritizing regional strength over head-to-head competition with frontier labs on general-purpose leaderboards. The ZDR/BYOK requirement also suggests enterprise and compliance-focused customers are a target segment.
Related Articles
Ant Group releases Ling-3.2-flash: 560B-parameter multimodal model with 1M-token context
Ant Group released Ling-3.2-flash on November 5, a native multimodal model that accepts text, images, and video. It has 560 billion total parameters, activates 25 billion per inference step, and supports a one-million-token context window.
Thinking Machines Lab launches Inkling 2 Small, a 594B-parameter sparse MoE model, on Tinker
Thinking Machines Lab introduced Inkling 2 Small on November 5, 2026. The sparse Mixture-of-Experts model has 594 billion total parameters and 17 billion active parameters, and is available through Tinker with closed weights at launch.
Baidu open-sources ERNIE 5.5 Flash Base, an MoE model with 1M-token context for fine-tuning and inference
Baidu open-sourced ERNIE 5.5 Flash Base on November 5. It is a Mixture-of-Experts base model pretrained on 33.5 trillion tokens, with 128K pretraining context and up to 1M tokens in fine-tuning and inference. Baidu claims it sets state-of-the-art results among models of the same size.
NVIDIA GR00T 2 reaches 40% zero-shot success on 50 unseen robot tasks using human video
NVIDIA introduced GR00T 2 on November 9, 2026, at the Conference on Robot Learning (CoRL) in Austin. The company reported 40% zero-shot success on 50 tasks absent from the target robot's training data for a model trained partly on human video.
Comments
Loading...