Nous Research released Hermes Agent on June 16, 2026. The launch includes two open-weight models, an 8B and a 70B, both under the Apache 2.0 license. Each model has a 131,072 token context window and is built for tool calling, browser use, and multi-step agent tasks. The files are available through Hugging Face and GitHub, but no official hosted API exists at launch. This is a local-first release aimed at developers who want a free agent without per-token billing. If you are tracking the best zero cost model options, see our list of free AI models in 2026.
Nous Research is the team behind the Hermes fine-tunes that have been popular in the open-source community for years. The vendor announcement on Nous Research positions Hermes Agent as a dedicated agent model rather than a general chat model. The release lands on Hugging Face and GitHub as open weights plus inference scripts. Nous did not publish a single model card path in the announcement, so users should start from the vendor homepage or the official organization pages on Hugging Face. The 8B model is aimed at testing and edge use. The 70B model targets heavier automation on a single GPU or a modest local server.
Why this matters is straightforward. Closed agent products from OpenAI, Anthropic, and Google have moved toward usage-based billing, credit pools, and tighter free tiers in June 2026. An Apache 2.0 agent model with a 131k context window gives teams a way to run automation without watching a meter. The license permits modification, commercial use, and private deployment. Hermes Agent does not match the top closed benchmark scores, but it removes the pricing anxiety and the data exposure. The launch arrives during a period when free tier access is shrinking, as covered in our tracker on AI pricing changes.
The timing is important. On June 16, 2026, many developers were already dealing with Anthropic ending agent subsidies and OpenAI Codex moving to metered access. Hermes Agent gives those developers an off-ramp. The release includes benchmark numbers that show a real gap to proprietary agents but not an unusable one. You can run the 8B version on a 16GB laptop or the 70B on a 24GB GPU with quantization. This release does not solve the hardware problem, but it removes the subscription and per-token problem. The next sections compare the model family against three paid or limited free alternatives and explain the hardware costs that free licensing does not erase.
How Do the Top Options Compare?
| Agent | Best For | License | Context Window | AgentBench Score | Cost |
|---|---|---|---|---|---|
| Hermes Agent 70B | Local complex tasks | Apache 2.0 | 131,072 tokens | 66.2 | Free |
| Hermes Agent 8B | Low-resource testing | Apache 2.0 | 131,072 tokens | 54.8 | Free |
| OpenAI Codex Agent | Managed coding | Proprietary | 200,000 tokens | 72.1 | Usage-based |
| Anthropic Claude Agent | Enterprise tool calling | Proprietary | 200,000 tokens | 74.0 | Credit pool |
| Mistral Le Chat Agent | Lightweight hosted | Apache 2.0 (weights) | 128,000 tokens | 58.3 | Free tier plus paid |
Benchmark scores are public or self-reported as of June 2026 and vary by evaluation set. Hardware requirements assume 4-bit or 8-bit quantization for local models.
1. Hermes Agent 70B , Best for complex local agent tasks
The 70B version is the flagship Hermes Agent release. It has 70 billion parameters and a 131,072 token context window. Nous Research reports an AgentBench score of 66.2, which is useful but below the closed frontier agents from OpenAI and Anthropic. The model runs best on a 24GB or 40GB GPU using 4-bit or 8-bit quantization. Full precision requires around 140GB of VRAM, which makes quantization the practical default. The weights are released under Apache 2.0, so self-hosting teams can deploy it internally without license review.
The real advantage is cost control. A local 70B agent does not generate per-token fees. You pay for electricity and hardware, not for every tool call. That matters in 2026 because agent workloads can multiply tokens fast. We covered the agentic AI billing crisis for free users and how hidden token costs hurt teams. Hermes Agent 70B sidesteps that meter entirely when you run it on your own box.
The tradeoff is hardware and throughput. On a single RTX 4090 or A6000, the model may produce 10 to 15 tokens per second with 4-bit quantization. That is enough for background automation but too slow for interactive coding. You also need to manage prompts, memory, and tool schemas yourself. There is no official managed endpoint from Nous Research at launch. For teams that want a no-cost local agent and already own a capable GPU, this is the strongest open option announced in June 2026. This release does not remove hardware costs, but it removes the meter.
Key strengths:
- ✅ Runs locally with no per-token fees
- ✅ Apache 2.0 allows commercial and private use
- ✅ 131,072 token context supports long agent traces
- ✅ Strong local AgentBench score of 66.2
- ✅ Weights are open for fine-tuning
- ❌ Needs 24GB to 40GB of VRAM even with quantization
- ❌ No hosted API from Nous Research at launch
- ❌ AgentBench still trails closed rivals by several points
Who it’s for: Developers with a 24GB or larger GPU who want a private, free agent for long-running automation.
2. Hermes Agent 8B , Best for low-resource local testing
The 8B version is the lighter sibling. It has 8 billion parameters and the same 131,072 token context window. Nous Research reports an AgentBench score of 54.8 for the 8B. That is much lower than the 70B, but the model is designed to run where the 70B cannot. An 8B model in 4-bit form fits in about 6GB of memory. That makes it viable on a laptop with 16GB of RAM and no discrete GPU.
You can use the 8B for prompt testing, tool schema iteration, and lightweight browser tasks before you move to the 70B. The same Apache 2.0 license applies. There is no separate commercial restriction for the smaller model. The release is hosted on Hugging Face under the Nous Research organization, but you should start from the Nous homepage to confirm the current file names. Free tier limits across the industry are getting stricter, as we covered in this June 2026 policy recap.
The main limitation is accuracy. The 8B model will fail on complex multi-step tasks that the 70B or a closed model can handle. It is not a replacement for Codex or Claude when precision matters. But it is a good smoke test for local tool calling and a cheap way to learn how an open agent behaves. If you have no GPU, the 8B is the realistic starting point for a completely free setup.
Key strengths:
- ✅ Runs on a 16GB laptop with CPU or integrated GPU
- ✅ Small download size and fast iteration
- ✅ Same Apache 2.0 license as the 70B
- ✅ Good for prompt and tool schema testing
- ❌ AgentBench score is low at 54.8
- ❌ Fails on complex multi-step tasks
- ❌ Still requires some setup for tool integration
Who it’s for: Developers with limited hardware who want to test a local open-source agent before investing in a GPU.
3. OpenAI Codex Agent , Best for managed cloud agent coding
OpenAI Codex Agent is the managed cloud alternative. It uses a proprietary model and a 200,000 token context window. OpenAI reports a higher AgentBench score around 72.1. The tool is strong at code generation, shell commands, and file edits. However, the pricing moved to usage-based billing in 2026. You pay for input and output tokens, with multiplier fees for agentic loops. That can surprise teams with a heavy automation workload.
The free tier for Codex Agent has tightened. We detailed the changes in ChatGPT Codex free tier and agentic coding. Closed pricing means you do not see the cost until the invoice arrives. The model itself is better than Hermes Agent 70B on most coding benchmarks, but the difference may not justify the cost for teams that run continuous background tasks. If you need managed infrastructure and top accuracy, Codex is the pragmatic choice.
The vendor homepage is OpenAI. The key tradeoff is lock-in. You cannot download the model or inspect the weights. Your prompts, code, and tool outputs flow through OpenAI servers. For data sensitive work, that is a real limitation. Hermes Agent offers a local, open-weight path that avoids that, but you give up managed convenience and some accuracy.
Key strengths:
- ✅ Higher benchmark scores for coding tasks
- ✅ Managed API with no local GPU requirement
- ✅ Strong ecosystem and tool integration
- ✅ 200,000 token context window
- ❌ Usage-based pricing with agentic multipliers
- ❌ Proprietary model with no local weights
- ❌ Data flows through third-party servers
Who it’s for: Teams that need managed, high-accuracy coding agents and can tolerate variable token costs.
4. Anthropic Claude Agent , Best for enterprise tool calling
Anthropic Claude Agent is the enterprise favorite for tool calling and long instructions. It uses a proprietary Claude model with a 200,000 token context window. Its AgentBench score is around 74.0, higher than Hermes Agent 70B. In June 2026 Anthropic replaced flat-rate agent access with a credit pool. Agent calls now drain credits based on tokens and tool use. That change caused backlash from developers.
The model is strong, but the cost is less predictable than before. For an enterprise that already runs Claude, the switch to credits is manageable. For an individual developer, the credit pool can feel like a meter. Hermes Agent’s no-meter local approach is the main open alternative.
Anthropic offers no open weights for Claude Agent. You can only use it through the API or first-party apps. The Anthropic homepage has the latest pricing. Claude Agent has better safety tooling and longer context than the Hermes 70B, but that comes with vendor control. If compliance requires an audit trail on Anthropic infrastructure, Claude remains a top choice.
Key strengths:
- ✅ Strong tool calling and long context
- ✅ Enterprise safety and permission controls
- ✅ Managed API with no local hardware
- ✅ Better AgentBench score than local models
- ❌ Credit pool pricing is unpredictable
- ❌ Closed weights and vendor lock-in
- ❌ No true free unlimited tier
Who it’s for: Enterprises that need managed tool calling, auditability, and top accuracy with a budget for usage-based spend.
5. Mistral Le Chat Agent , Best for a lightweight hosted free tier
Mistral AI offers Le Chat Agent as a hosted option with a free tier. The underlying weights are open for many Mistral models, but the managed service has limits. The context window is 128,000 tokens. Its AgentBench score is around 58.3, slightly above Hermes Agent 8B but below the 70B. The free tier gives casual users a way to try agent features without a GPU.
Mistral’s free tier is not unlimited. We compared the Mistral Le Chat free tier and its limits. Heavy agent workloads require a paid plan or an API key. The open-weight route is possible for some Mistral models, but the best agent-tuned weights are not always the ones served in the free chat. The Mistral AI homepage lists current models and pricing.
Compared to Hermes Agent, Mistral Le Chat is easier to start with because it is hosted. You do not need to set up a local runtime. But the free tier resets, rate limits, and data policies still apply. For a truly free and private agent, Hermes Agent 70B or 8B is the stronger open-source play. For a quick test of agent behavior, Mistral’s free tier is fine.
Key strengths:
- ✅ Hosted free tier with no local setup
- ✅ Open-weight ecosystem from a European vendor
- ✅ Decent 58.3 AgentBench for a lighter model
- ❌ Free tier has rate and feature limits
- ❌ Agent-tuned weights may differ from hosted model
- ❌ Less control than a fully local Apache 2.0 model
Who it’s for: Casual developers who want a quick hosted agent test without committing to local hardware or paid subscriptions.
Frequently Asked Questions
Is Hermes Agent actually free?
Yes. The models are free to download and use under the Apache 2.0 license. You still need hardware or cloud compute to run it. There is no per-token fee from Nous Research.
What license does Hermes Agent use?
Hermes Agent uses the Apache 2.0 license. You can modify, distribute, and use it commercially. You must include the license notice in distributions.
Can I run Hermes Agent on a laptop?
The 8B version can run on a 16GB laptop using 4-bit quantization. The 70B version needs about 40GB of memory with quantization. A discrete GPU is not required for the 8B, but it helps.
How does Hermes Agent compare to OpenAI Codex Agent?
OpenAI Codex Agent has a higher AgentBench score around 72.1, while Hermes Agent 70B is around 66.2. Codex is managed and usage-based. Hermes Agent is local and free but slower on modest hardware.
Where can I download Hermes Agent?
Start at the Nous Research homepage or the official Nous Research organization pages on Hugging Face and GitHub. The announcement did not publish a single model card path, so confirm file names from those official sources.
What context length does Hermes Agent support?
Both the 8B and 70B models support a 131,072 token context window. That is enough for long multi-step agent traces without constant truncation.
What Should You Remember?
- Release: NousResearch launched Hermes Agent on June 16, 2026 as an open-source local agent.
- Models: Two variants ship, an 8B and a 70B, both with 131,072 token context.
- License: Apache 2.0 permits commercial use, modification, and private deployment.
- Benchmarks: The 70B scores 66.2 on AgentBench, trailing closed agents but usable.
- Hardware: The 8B fits in 16GB RAM; the 70B needs about 40GB with quantization.
- Cost: Self-hosting removes per-token fees but not hardware and electricity costs.
- Privacy: Local execution keeps prompts, tool calls, and outputs on your own machine.
Free AI News is an independent editorial publication. Information about AI pricing, free-tier limits, and features changes frequently and may become outdated. Always verify current details through the vendor’s official pages. Affiliate links may earn a commission at no cost to you, and never affect our reporting.
