One open model dropped on June 12. The question is how you run it.

What Kimi K2.7-Code is, in the numbers that matter

Moonshot AI shipped K2.7-Code on June 12, 2026 as a coding-first model built on the K2.6 architecture, with open weights and a permissive license on its model card. The figures below come from that card and Moonshot's release notes. Read the last one as a vendor claim, not a settled fact: a +21.8% jump on a benchmark the vendor designed and runs itself is a reason to test, not a reason to believe.

Open weights at that scale are not something most people will run locally; the full model is hundreds of gigabytes. In practice you will reach it through a hosted provider or a gateway. Which is exactly where the next part comes in: the gateway is also how it reaches your agent loop.

The gap every roundup leaves open

Search this topic and you get model cards, spec tables, and pricing comparisons. All of it ends at the same place: the model exists, here are its numbers. The thing nobody answers is the practical one. You use an AI coding agent every day, with sessions you do not want to lose, a forking habit, and a long context you do not want silently truncated. A new open model drops. How do you actually try it inside that setup, today, without throwing away your session machinery and learning a new app?

That gap is the spine of this page. The answer on macOS is a single override, and it is small enough to be unglamorous, which is probably why no roundup bothers with it.

The one setting: Fazm overrides ANTHROPIC_BASE_URL, nothing else

Fazm is a native macOS app that wraps Claude Code, Codex, and Gemini as agent-loop backends through the Agent Client Protocol. Its Custom API Endpoint setting does exactly one thing, and the source is blunt about it: it overrides ANTHROPIC_BASE_URL. When the endpoint is on, Claude-compatible requests route to your gateway instead of Fazm's built-in credits. Put a server that speaks the Anthropic Messages API in front of an open model, and the same persistent Mac UI now drives that model.

The sharp edge is in the third and fourth lines above. Because the override only affects the Claude path, the model you select in the picker has to be a Claude id for traffic to reach your endpoint. Pick Codex or Gemini while the endpoint is set and your requests quietly bypass it. Fazm guards against that silent failure with a warning that lives in Desktop/Sources/MainWindow/Pages/SettingsPage.swift: "The custom endpoint only applies to Claude models. Switch to a Claude model for your requests to reach it." That guard was added in a June 15, 2026 commit titled "Add warning for non-Claude models in custom endpoint settings."

Wiring a day-old open model into the loop, in four steps

This is the whole procedure. The first step is the only one that takes real work, and you do it once; after that, every future release is the same four steps with a different model loaded behind the gateway.

What the endpoint is allowed to proxy to

The in-source comments name the supported shapes directly. One comment in ACPBridge.swift describes the endpoint as allowing proxying through Copilot and corporate gateways. The error handling a few lines down names LM Studio and Ollama as local servers, including a message that tells you to load a model in your local server when the endpoint reports nothing loaded. The settings search entry for the feature carries the keywords endpoint, proxy, base url, anthropic, copilot, gateway, and corporate.

So three concrete routes exist for a new open model: a hosted Claude-compatible gateway (a service that re-exposes the model behind the Anthropic wire format), a corporate proxy your team already runs, or a local server on your own machine. In every case Fazm does not load the weights; it sends the agent loop to wherever your endpoint points.

Why this is worth doing instead of opening another app

The reason to route a new model through the endpoint rather than downloading a separate chat client is that the model is the cheap part to swap. What is expensive to give up is the session layer sitting above it. In Fazm, that layer is the same regardless of which model answers: chats survive a Mac restart and every window auto-restores, forking a conversation is one click into a new window that carries the full prior context while leaving the original untouched, and the history stays live with no auto-compacting for the lifetime of the window.

Try a day-old release the usual way and you start a brand-new conversation in a brand-new app, with none of that. Try it through the endpoint and you are evaluating the model under the exact working conditions you care about, with a real project loaded and a real session you can fork to compare against your current model. That is a fairer test, and it is faster.

How to verify every claim on this page

1. For the model: open moonshotai/Kimi-K2.7-Code and read the card for the release date, the parameter and context figures, and the license. The commit history on the repo carries the June 12, 2026 timestamps.
2. For the endpoint behavior: open github.com/mediar-ai/fazm, read Desktop/Sources/MainWindow/Pages/SettingsPage.swift for the ANTHROPIC_BASE_URL note and the non-Claude warning string, and Desktop/Sources/Chat/ACPBridge.swift for the Copilot, corporate, LM Studio, and Ollama references.
3. For the broader picture: open huggingface.co/papers and github.com/trending. The trending order is volatile, but the submission dates are stable.