How It Works
When you set up Hermes Agent with the Nous Portal, you get a single OAuth subscription that proxies 300+ frontier models through OpenRouter under the hood. No separate API keys, no zero-balance errors, and mid-session model switching via /model — all billed against one subscription.
The models are organized into 14 model families, each with different strengths for agentic workflows. Nous Research categorizes them into tiers based on performance in reasoning, coding, tool use, and cost efficiency.
The same Portal subscription unlocks the Nous Tool Gateway, which routes Hermes Agent's tool calls through Nous-managed infrastructure — web search via Firecrawl, image generation via FAL (9 models under one endpoint), TTS via OpenAI, browser automation via Browser Use, and cloud terminal sandboxes via Modal. No separate accounts needed.
Frontier Tier — Maximum Reasoning & Agentic Performance
Use these for complex multi-step reasoning, deep analysis, and mission-critical agent tasks where accuracy matters most.
- Opus 4.7
- Opus 4.6
- Sonnet 4.6
- Haiku 4.5
Claude remains the gold standard for agentic reasoning. Claude Fable 5 (June 9, 2026) is the newest flagship — a quantum leap beyond Opus 4.8. Since Claude 3, the tiered lineup (Haiku = fast, Sonnet = balanced, Opus = maximum capability) has proven extremely effective for agent workflows.
- Constitutional AI training for safe, aligned tool use
- 1M+ token context window on Opus
- Best-in-class long-context reasoning for agent memory
- Strongest tool-use and function-calling benchmarks
- Fable 5: newly released June 9, 2026 — state-of-the-art across all benchmarks
- GPT-5.5
- GPT-5.5 Pro
- GPT-5.4 Mini
- GPT-5.4 Nano
- GPT-5.3 Codex
GPT-5.5 and GPT-5.5 Pro represent the latest generation of OpenAI's frontier models. GPT-5.3 Codex remains the preferred variant for code-heavy agentic tasks, with strong integration into the broader OpenAI ecosystem.
- GPT-5.5 Pro: maximum reasoning capability for complex agents
- GPT-5.3 Codex: strongest coding model in the family
- GPT-5.4 Mini: best cost-performance balance for high-volume agents
- GPT-5.4 Nano: lightweight variant for constrained environments
- Extensive API ecosystem and third-party integrations
- Gemini 3 Pro Preview
- Gemini 3 Flash Preview
- Gemini 3.1 Pro Preview
- Gemini 3.1 Flash Lite Preview
Gemini 3.1 Pro, 3 Deep Think, 3.5 Flash, and 3.1 Flash Lite were released May 19, 2026. Gemini's multilingual strength (30+ languages) and native multimodal capabilities make it ideal for agents that need to process text, images, audio, and video. The Flash variants offer exceptional speed for high-throughput agent tasks.
- Best multilingual model family — 30+ languages natively
- Gemini 3.1 Pro: strong reasoning with native image/video/audio processing
- Gemini 3.5 Flash: fastest variant for high-volume tasks
- Deep Think mode for complex multi-step reasoning
- Best-in-class long-context window (up to 2M tokens)
- DeepSeek V4 Pro
Founded by Liang Wenfeng in Hangzhou, DeepSeek disrupted the industry in early 2025 with V3 and R1 models that matched frontier performance at a fraction of the cost. DeepSeek V4 Pro is the latest flagship. The company's open-weight approach and training framework innovations have been widely cited.
- DeepSeek V4 Pro: latest flagship with strong reasoning and coding
- MoE (Mixture of Experts) architecture for efficient inference
- Strong performance on coding and math benchmarks
- Cost-efficient alternative to Western frontier models
Fast Tier — High-Throughput Agentic Work
Use these for routine agent tasks, fast tool calls, and high-volume workflows where speed matters more than absolute reasoning depth.
- Qwen3.7-Max
- Qwen3.6-35B-A3B
Qwen3.7 Max (May 18, 2026) and Qwen3.7 Plus are the latest from Alibaba's Qwen family, which includes both proprietary and open-weight models. Qwen3.6-35B-A3B (April 15, 2026) is a sparse MoE model — 35B total parameters with only 3B active — making it extremely efficient for agent deployment.
- Qwen3.7 Max: strongest model in the family, competitive with Claude/GPT
- Qwen3.6-35B-A3B: 35B params / 3B active — ultra-efficient MoE
- Strong multilingual support (Chinese, English, and more)
- Open-weight variants available for self-hosting
- Excellent coding and math capabilities
- Kimi K2.6
Moonshot AI founded in March 2023, released Kimi in October 2023. Originally famous for 128K context support, Kimi K2 (July 2025) showed strong coding benchmarks with MIT license. Kimi K2.7 Code (June 16, 2026) is the latest coding-focused variant. Kimi K2 uses a modified MIT License.
- Kimi K2.6: strong reasoning and coding for agentic tasks
- Moonshot's proprietary models optimized for Chinese/English bilingual tasks
- Long-context processing heritage (128K+ tokens)
- Kimi K2.7 Code: newest variant, June 16, 2026
- GLM-5.1
Zhipu AI (branded Z.ai internationally since 2025) was founded in 2019 by Tang Jie and Li Juanzi, with Zhang Peng as CEO. The company went public on SEHK:2513. GLM (General Language Model) is their flagship family. GLM-5.2 and GLM-5.1 are the latest releases, with GLM-4.7 and GLM-4.7 Flash also available.
- GLM-5.1: latest frontier model from Z.ai
- Strong Chinese-language understanding and generation
- 800+ employees (2024), publicly traded SEHK:2513
- GLM-4.7 Flash: fast variant for high-throughput agents
- MiniMax M2.7
Founded by Yan Junjie, Yang Bin, and Zhou Yucong in December 2021, MiniMax is now publicly traded on SEHK:100. Revenue grew from $30.5M (2024) to $79M (2025). Beyond LLMs, MiniMax also makes Hailuo (video generation), Speech (TTS), Music, and Talkie (chat). MiniMax M3.0 and Hailuo 2.3 are their latest products.
- MiniMax M2.7: strong general-purpose model
- Full AI ecosystem: LLMs, video, TTS, music, chat
- 415 employees (2025), SEHK:100 listed
- MiniMax Agent: built for agentic workflows
Regional Champions — Specialized Strengths
These models excel in specific domains or regions, offering unique capabilities that complement the major Western and Chinese families.
- Grok 4.3
Founded by Elon Musk in March 2023, xAI is now a subsidiary of SpaceX (Michael Nicolls as president). Grok 4.3 Beta (April 17, 2026) is the latest, with Grok 4.1 Fast (November 2025) as the high-speed variant. Grok has unique real-time access to X (formerly Twitter) data. Revenue: $3.2B (2025), 1,200+ employees.
- Grok 4.3 Beta: latest frontier model from xAI
- Grok 4.1 Fast: high-speed variant for agent workflows
- Real-time X/Twitter data access (unique differentiator)
- Built into Tesla OS, Android, iOS, ChromeOS
- Grok 1 released under Apache-2.0; later models under xAI Community License
- Nemotron-3 Super 120B-A12B
Founded by Jensen Huang, Chris Malachowsky, and Curtis Priem in 1993, NVIDIA is the dominant GPU manufacturer and a major player in AI infrastructure. Nemotron-3 Super 120B-A12B is a 120B-parameter MoE model with 12B active parameters — designed for efficient evaluation and reward modeling.
- Nemotron-3 Super: 120B params / 12B active — ultra-efficient MoE
- Specialized in evaluation and reward modeling for agent feedback
- NVIDIA dominates AI infrastructure (CUDA ecosystem)
- Strong open-source contributions to AI evaluation
- Hunyuan 3 Preview
Founded by Pony Ma in 1998, Tencent is one of the world's largest technology companies (SEHK:700, Hang Seng Index component). Beyond gaming and social media (WeChat), Tencent has invested heavily in AI. Hunyuan 3 Preview is their latest large language model.
- Hunyuan 3 Preview: latest Tencent LLM
- Massive Chinese-language training corpus from WeChat ecosystem
- Strong in gaming AI, content generation, and search
- Integrated with Tencent Cloud infrastructure
- MiMo V2.5 Pro
Founded by Lei Jun in 2010, Xiaomi (SEHK:1810) is one of the world's largest smartphone manufacturers and a growing player in AI. Xiaomi has expanded into electric vehicles (SU7), smart home, and now generative AI. MiMo V2.5 Pro is their latest AI language model, designed for on-device and cloud deployment across their ecosystem.
- MiMo V2.5 Pro: latest Xiaomi AI language model
- Optimized for on-device deployment (phones, smart home, EVs)
- Strong integration with Xiaomi's hardware ecosystem
- Growing investment in large-scale AI models
- Step 3.5 Flash
Founded by former Microsoft employees Jiang Daxin, Zhu Yibo, and Jiao Binxing in April 2023, StepFun has been dubbed one of China's "AI Tiger" companies. Investors include Tencent, Qiming Venture Partners, and Shanghai State-owned Capital Investment. In July 2025, StepFun announced the "Model-Chip Ecosystem Innovation Alliance" with Huawei, Biren Technology, Moore Threads, and Enflame.
- Step 3.5 Flash: fast variant for high-throughput agents
- "AI Tiger" company with strong VC backing
- Model-Chip ecosystem alliance with Chinese chip makers
- Former Microsoft team with strong engineering pedigree
Open Source Homegrown — Built for Hermes
- Hermes-4-70B
- Hermes-4-405B (chat)
Hermes is Nous Research's own model family, purpose-built for agentic workflows and the Hermes Agent platform. Named after the Greek messenger god, these models are optimized for tool use, multi-step reasoning, and reliable agent behavior. Hermes-4-405B supports chat mode (see note below in the Portal docs), making it suitable for conversational agents alongside its core agentic capabilities.
- Hermes-4-70B: efficient open-weight model for self-hosted agents
- Hermes-4-405B: maximum capability model (chat mode available)
- Optimized for tool use, function calling, and multi-step agents
- Natively integrated with Hermes Agent — first-party support
- Open weights available for full transparency and self-hosting
Quick Reference — Model Selection by Use Case
| Use Case | Primary Pick | Alternative | Tier |
|---|---|---|---|
| Maximum Reasoning | Claude Opus 4.7 | GPT-5.5 Pro | Frontier |
| Best Value Reasoning | DeepSeek V4 Pro | Qwen3.7-Max | Frontier |
| Coding Agent | GPT-5.3 Codex | Qwen3.6-35B-A3B | Fast |
| High-Throughput Tasks | Gemini 3.1 Flash Lite | Haiku 4.5 | Fast |
| Long Context | Claude Opus 4.7 | Gemini 3 Pro | Frontier |
| Chinese Language | Qwen3.7-Max | GLM-5.1 | Fast / Regional |
| Multimodal (Image/Video) | Gemini 3 Pro | MiniMax M2.7 | Fast |
| Real-Time Data Access | Grok 4.3 | — | Regional |
| Self-Hosted / On-Device | Hermes-4-70B | Qwen3.6-35B-A3B | Open Source |
| Agent Evaluation | Nemotron-3 Super | — | Regional |
| On-Device / IoT | MiMo V2.5 Pro | — | Regional |
| Native Hermes Integration | Hermes-4-405B | Hermes-4-70B | Open Source |
Configuration Tips
Model Switching Mid-Session
Use /model mid-session to switch between models. For example, switch to Claude Sonnet 4.6 for code review, then to Gemini 3 Pro for long-context document analysis — all under one subscription, no new credentials needed.
Cost Optimization Strategy
Primary agent: Claude Sonnet 4.6 or GPT-5.4 Mini for routine tasks
Complex reasoning: Claude Opus 4.7 or GPT-5.5 Pro when accuracy is critical
High-volume: Gemini 3.1 Flash Lite or Haiku 4.5 for fast, cheap iterations
Self-hosted: Hermes-4-70B or Qwen3.6-35B-A3B for local deployment
The Portal Setup
The fastest way to get started:
hermes setup portalThis single command runs the Portal OAuth, lets you pick a Nous model, sets Nous as your inference provider in config.yaml, and turns on the Tool Gateway. You're ready to hermes chat immediately after.
Don't have a subscription? Visit portal.nousresearch.com/manage-subscription — sign up, then come back and run the setup command above.
Routing Behavior
Routing happens through OpenRouter under the hood, so model availability and failover behavior matches what you'd get with an OpenRouter key — just billed against your Nous subscription instead. You can also enable specific gateway tools (e.g., web search but not image generation) and mix the gateway with your own backends.
280+ Additional Models
Beyond the 14 families covered here, the Nous Portal includes 280+ additional models — the full agentic frontier. This includes specialized models for:
- Specialized reasoning — math, law, medicine, science
- Code generation — language-specific (Python, Rust, Go, etc.)
- Multimodal — image generation, video, audio, 3D
- Edge / mobile — on-device models for phones and IoT
- Emerging families — new entrants and academic models
The catalog grows regularly as new models are added to the OpenRouter integration. Check portal.nousresearch.com for the latest additions.
Summary
Nous Research's Hermes Agent Creator provides access to the most comprehensive model catalog available — 300+ frontier models across 14 families, all billed through a single subscription with OpenRouter routing. The key takeaways:
- For maximum reasoning: Claude Opus 4.7, GPT-5.5 Pro
- For cost-efficient frontier: DeepSeek V4 Pro, Qwen3.7-Max
- For high-throughput: Gemini 3.1 Flash Lite, Haiku 4.5
- For coding: GPT-5.3 Codex, Qwen3.6-35B-A3B
- For Chinese: Qwen3.7-Max, GLM-5.1, Kimi K2.6
- For self-hosted: Hermes-4-70B, Qwen3.6-35B-A3B
- For native Hermes: Hermes-4-405B / Hermes-4-70B