Upstage Launches Solar Mini 4, an Agent-Optimized Language Model
Upstage announced Solar Mini 4, an LLM that activates 3 billion of 35 billion parameters per token, reporting benchmark scores and internal cost comparisons for agent workloads.
Quick answer
What did Upstage announce about Solar Mini 4?
Upstage launched Solar Mini 4, an agent-optimized language model with 35 billion total parameters that activates only 3 billion per token. Upstage reported a 24.1 score on the Artificial Analysis Intelligence Index, a 512,000-token context window, and API pricing starting at $0.10 per million input tokens.
Key takeaways
- Upstage launched Solar Mini 4, an agent-optimized model with 35 billion total parameters that activates only 3 billion per token.
- Upstage says the model scored 24.1 on the Artificial Analysis Intelligence Index v4.3.2, which it describes as the highest score among models with 3 billion active parameters in that comparison.
- The model supports a 512,000-token context window, tool calling, parallel tool calling, and structured JSON outputs, according to Upstage.
- Standard API pricing is $0.10 per million input tokens, $0.01 per million cached input tokens, and $0.40 per million output tokens, Upstage says.
- In an internal cost comparison Upstage ran against MiMo-V2.5, Solar Mini 4 completed 1,000 sets of agent tasks for an estimated $1.25 versus $2.20 for MiMo-V2.5.
What Upstage announced
Upstage has launched Solar Mini 4, a new large language model the company says was pretrained from the ground up and optimized for agent workloads. According to Upstage, the model uses a mixture-of-experts architecture with 35 billion total parameters but activates only 3 billion of them per token, a design intended to lower compute costs for repetitive agent tasks such as information retrieval, structured output generation, and tool use.
Benchmark results Upstage reported
Upstage says Solar Mini 4 scored 24.1 on the Artificial Analysis Intelligence Index v4.3.2, which the company describes as the highest result among models with 3 billion active parameters included in that comparison. Upstage says this places it ahead of Qwen3.6 35B A3B, which scored 18, and Nemotron 3.5 Lightning, which scored 13.
The company also compared Solar Mini 4 to models with more active parameters. Upstage says Solar Mini 4 activates roughly one-tenth as many parameters as Gemma 4 31B while scoring 9.1 points higher on the AA Index, and that it scored 1.1 points higher than Nemotron 3 Ultra, which has 55 billion active parameters. Upstage says Solar Mini 4's score was 0.9 points lower than MiMo-V2.5 but delivered what it called comparable performance using one-fifth as many active parameters.
On other evaluations, Upstage reported that Solar Mini 4 scored 83.3% on AA-LCR, a long-context reasoning benchmark; 47.6% on SciCode, a code-based scientific problem-solving benchmark; and 25.8% on Humanity's Last Exam.
For agent-specific testing, Upstage says the model scored 22.3% on AutomationBench-AA, which evaluates multi-step workflows across real SaaS applications, and 47.2 on a separate benchmark called τ³-Banking, which tests policy application and tool use across multi-turn interactions.
Specifications and pricing
Upstage says Solar Mini 4 supports a context window of up to 512,000 tokens and up to 128,000 output tokens. The model supports tool calling, parallel tool calling, JSON mode, and JSON Schema-based structured outputs. Upstage says reasoning is not enabled by default but developers can adjust a reasoning_effort setting for tasks that require multi-step analysis.
Standard API pricing, according to Upstage, is $0.10 per million input tokens, $0.01 per million cached input tokens, and $0.40 per million output tokens. Upstage says Console pricing carries a 70% launch discount through October 10 UTC, after which standard pricing takes effect.
Upstage says the model is available through the Upstage Console API, Solar Chat, and OpenRouter, with on-premises deployment also supported, and that it will be free to use on Hermes Agent for a limited time starting October 5. The company says a quantized version of the model can run on a single H100 80GB GPU for on-premises deployment, and that in internal testing, Solar Mini 4 sustained more than 70 tokens per second per request while processing 32 concurrent requests on two H100 GPUs.
Cost comparison against MiMo-V2.5
Upstage ran an internal test comparing Solar Mini 4 to MiMo-V2.5 on three agent tasks, including code analysis and review and calendar coordination across multiple tools, each run three times per model. Upstage says both models passed all quality checkpoints, and that the estimated cost of completing 1,000 sets of the three tasks was $1.25 with Solar Mini 4 versus $2.20 with MiMo-V2.5. For a multi-step tool-calling task involving checking a calendar and sending messages to attendees, Upstage reported an estimated cost of $0.36 per 1,000 runs with Solar Mini 4 compared with $0.61 with MiMo-V2.5. Upstage noted that MiMo-V2.5 was accessed through a third-party router in this test, and that latency was excluded from the comparison.
Source: Upstage, "Solar Mini 4: Built for High-Volume Agent Workloads," published October 1, 2026.
Frequently asked questions
- What is Solar Mini 4?
- Solar Mini 4 is a large language model from Upstage, pretrained from the ground up and optimized for high-volume agent workloads, according to the company.
- How many parameters does Solar Mini 4 use?
- Upstage says the model has 35 billion total parameters but activates only 3 billion of them per token, using a mixture-of-experts architecture.
- What does Solar Mini 4 cost to use?
- Upstage lists standard API pricing at $0.10 per million input tokens, $0.01 per million cached input tokens, and $0.40 per million output tokens.
- Where is Solar Mini 4 available?
- Upstage says the model is available through the Upstage Console API, Solar Chat, and OpenRouter, with on-premises deployment also supported, and will be free on Hermes Agent for a limited time starting October 5.