Solar Mini 4 Brings 3B MoE and 512K Context to Nous Portal

Upstage’s sparse architecture balances low inference overhead with an expansive context window, backed by two weeks of free developer access.

Nous Research added Solar Mini 4 to the Nous Portal catalog on Monday (October 5, 2026), granting two weeks of complimentary access. Built by South Korea’s Upstage, the model adopts a sparse design to tackle a common pain point in autonomous AI agents: reading extensive histories without running up prohibitive compute bills on every operational step. Notably, Nous Research has an obvious commercial interest in promoting the system, as it operates the portal and seeks to draw developers into its ecosystem, including the Hermes Agent framework.

First deployed by Upstage in late September 2026 as snapshot solar-mini4-260922, the model targets workflows where an AI repeatedly invokes tools to complete complex tasks. In those operational cycles, latency and the cost per API call matter just as much as theoretical benchmark scores.

What the Solar Mini 4 MoE Architecture Means

The MoE acronym stands for Mixture of Experts. Instead of running every input token through the entire neural network, the system acts like a specialized team where only the relevant experts fire for any given word. Solar Mini 4 houses 35 billion total parameters, yet activates only 3 billion per token. The library analogy fits neatly: you keep an entire 35-volume reference set on hand, but you only open three volumes to answer a quick inquiry.

The primary practical feature is a 512K input context window (exactly 524,288 tokens), paired with a generation ceiling of up to 128K tokens (exactly 131,072 tokens). In practical terms, 512K tokens can hold well over a thousand pages of technical documentation or several weeks of multi-turn agent logs in active memory at once.

Benchmarks and Focus on Autonomous Agents

Upstage built the system specifically for external software execution. The engine supports parallel function execution (parallel_tool_calls=True), formatted JSON schema outputs, and the company’s Solar Decide protocol, which accelerates single-step routing decisions.

On benchmark ladders, Solar Mini 4 registered 24.1 points on the Artificial Analysis Intelligence Index (v4.3.2), leading all evaluated models in the 3B active parameter tier. In documentation published by Upstage, the model outscored larger dense architectures, finishing 9.1 points ahead of Gemma 4 31B and 1.1 points above Nemotron 3 Ultra 55B. For automated workflows, the company reported 22.3% on AutomationBench-AA and 47.2 points on the τ³-Banking benchmark.

Costs, Caveats, and What Remains Unproven

Unlike prior Upstage releases that provided open weights for direct download, Solar Mini 4 remains strictly proprietary. Developers can only access the model through commercial cloud APIs or authorized corporate on-premises deployments.

Outside the promotional window, Upstage lists standard pricing at $0.10 per million input tokens ($0.01 for prefix caching, which reuses repetitive system instructions) and $0.40 per million output tokens. On Nous Portal, the posted schedule sets pricing at $0.05 for input and $0.20 for output per million tokens.

Real-world deployments require some caution. An independent technical breakdown by MyClaw.ai pointed out that factual recall across the full 512K window without middle-context degradation (the classic lost-in-the-middle challenge) lacks broad third-party verification, making bulletproof retention across long contexts a technical RUMOR for now. Post-promotional rate limits and billing structures on the portal also remain subject to future confirmation.