Open-Source AI Agent Roundup, September 2026: Codex Harness, SoL-Pi, MiMo-V2.6 and Kimi K3

In the last two months, open source has closed much of the gap on both halves of the agent stack. OpenAI open-sourced the Codex harness under Apache-2.0, and new open-weight models from Xiaomi and Moonshot reached the trillion-parameter class. Here is what shipped, under which license, and what you can actually build with it.

Short answer: you can now build a production agent entirely from permissively licensed parts. Use an Apache-2.0 harness (OpenAI's Codex), an MIT efficiency layer (NVIDIA SoL-Pi) or assistant (Hermes Agent), and MIT-licensed open-weight models (Xiaomi MiMo-V2.6). The main catch is licensing. Not everything labelled "open" is open source, and Kimi K3 is the clearest example.

Roundup table

OpenAI open-sources the Codex harness

The biggest structural change is OpenAI publishing Codex as a reusable platform. The openai/codex repository is Apache-2.0 and gives you three ways in:

• codex exec runs a bounded agent task and returns structured output, which suits CI jobs and scripts.
• The Codex SDK starts, resumes and streams tasks from your own code.
• The Codex app-server holds persistent conversations with streamed events, interrupts, custom tools and approval requests, for embedding an agent inside your own product.

OpenAI draws a clear line: the harness is open, but model access and managed services are not. Because the harness is open, other teams can study and improve the loop that OpenAI says lifted GPT-5.6 Sol from 13.3% to 38.3% on ARC-AGI-3.

NVIDIA SoL-Pi: open-source token savings

SoL-Pi (MIT) sits on top of an unmodified Pi coding agent and adds four opt-in efficiency mechanisms. NVIDIA, NTU and MIT researchers found them by running auto-research loops over 535 executable environments. Reported results are up to 49% less token traffic and about 33% lower API cost. It installs with pi install git:github.com/NVlabs/SoL-Pi. The method may matter more than the package: an agent optimizing its own harness is now a published, reproducible technique.

Hermes Agent 0.21: the self-hosted assistant matures

Nous Research's MIT-licensed Hermes Agent shipped v0.21.0 "Pantheon" on August 31, with multi-agent Bot Mode, agent-to-agent messaging, scheduled jobs with memory, and browser automation. Four September patches followed, focused on stability. It's the closest open-source equivalent to a consumer agent like Meta Muse that you can run on your own hardware.

Xiaomi MiMo-V2.6: MIT-licensed, trillion-parameter class

Xiaomi released MiMo-V2.6 on September 22. The Pro and Flash checkpoints and a 9B Qwen distill are on Hugging Face under the MIT license. Xiaomi describes Pro as a trillion-parameter omni-modal reasoning model. The 9B distill runs on a single GPU, which makes it a practical local model for agent sub-tasks.

Kimi K3: open weights, custom license

Moonshot's Kimi K3 (July 26) is the largest open-weight model released so far, at 2.8 trillion parameters. Its license is custom, though. The Kimi K3 License requires model-as-a-service businesses above a revenue threshold to sign a separate agreement, and requires attribution in some commercial products. Read it before you build on it.

How to choose

• Embedding an agent in your SaaS: Codex app-server or SDK (Apache-2.0).
• Private team assistant on your own servers: Hermes Agent (MIT) with a local model.
• Cutting agent spend: SoL-Pi's techniques, even if you don't use Pi.
• Commercial product on open weights: prefer MIT or Apache-2.0 models like MiMo-V2.6 over custom licenses.

For the commercial CLIs, see agent harness releases for September 2026. For private deployment help, see our local AI swarm guide.

Frequently asked questions

Is the OpenAI Codex harness open source?

Yes. The openai/codex repository, which includes the Codex CLI, the app-server and the Codex SDK, is licensed under Apache-2.0. OpenAI says the open-source layer is the harness and integration surface, while model access and managed services remain separate.

What is Xiaomi MiMo-V2.6?

MiMo-V2.6 is Xiaomi's model family, released September 22, 2026. The Pro and Flash models and a 9B Qwen distill are published on Hugging Face under the MIT license. Xiaomi describes Pro as its trillion-parameter flagship reasoning model.

Is Kimi K3 open source?

Kimi K3's weights are downloadable on Hugging Face, but it uses a custom Kimi K3 License rather than a standard open-source license. Model-as-a-service businesses above a revenue threshold need a separate agreement with Moonshot AI, so check the license before commercial use.

What is the best open-source agent harness in 2026?

For embedding an agent in your own software, the Apache-2.0 Codex app-server and SDK are the most complete. For a self-hosted personal or team assistant, Hermes Agent (MIT) is the most active. For fully on-device agents, Stanford's OpenJarvis (Apache-2.0) is designed for local-first use.

What is the difference between open weights and open source?

Open weights means you can download and run the model. Open source, in the strict sense, also means a license that allows use, modification and redistribution without field-of-use restrictions. MIT and Apache-2.0 models meet that bar; custom licenses with revenue thresholds do not.