June 2026 field notes

Hugging Face new models or GitHub trending AI, June 2026: trending is not the same as runnable

If you opened the two feeds everyone watches this month, you saw a wall of new names. The question that actually matters is quieter: of the models trending on Hugging Face and the repos trending on GitHub, which can you put behind your agent today? That gap, and the single setting that closes it, is what this page is about.

Matthew Diakonov, Written with AI

Published June 22, 20267 min read

Direct answer · verified June 22, 2026

There is no official, dated June 2026 list of new AI on either site. Both rank discovery by a live, rolling score, so the real answer is two feeds you bookmark, not an article you read once:

Hugging Face: huggingface.co/models (sort by trending) and the trending papers tab.
GitHub: github.com/trending filtered to the day.

The catch most roundups skip: a name on either feed does not mean it runs in your coding agent yet. Read on for why, and the one field that fixes it.

A sample of what scrolled past on the trending tabs this month. Some are open weights you can download, some are API-only, some are repos, not models. That distinction is the whole point.

DeepSeek V4 (open, MIT)GLM series (open)Qwen 3.7 Max (API-only, no weights)Gemma 4 (open, Apache 2.0)Llama 4 (open)Kimi (open)deer-flow (repo)airllm (repo)firecrawl (repo)

Three states, not one

The mistake is treating a trending feed as a shopping list. A model or repo actually moves through three states, and only the last one matters for daily work:

State	What the feed tells you	What it does not tell you
Trending	People are clicking, starring, and citing it right now.	Whether weights exist, or it is API-only hype.
Downloadable	There is a weights file and a license you can read.	Whether your tools speak its API shape.
Runnable in your agent	Nothing. The feed stops caring here.	This is the only state that changes your week.

Qwen 3.7 Max is the clean example of the first trap: it trended on benchmark coverage while shipping API-only, with no open weights to download. DeepSeek V4 cleared the second state, real MIT-licensed weights on Hugging Face, but landed in the third trap, the API shape, which is where most people quietly give up.

The wall nobody mentions: format, not weights

Say you cleared the licensing hurdle and have an open June 2026 model serving locally. Out of the box it almost certainly exposes an OpenAI-compatible endpoint, because that is the vLLM and SGLang default. A coding agent built on Claude Code, like fazm, does not speak that. It speaks the Anthropic Messages API. Same idea, two incompatible wire formats:

Same conversation, two API shapes

# what most June 2026 open weights expose
# (vLLM / SGLang default, OpenAI-compatible)
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"deepseek-v4","messages":[...]}'
# path:  /v1/chat/completions
# schema: OpenAI messages array

0% fewer lines

This is why a trending name does not drop into your workflow. You are not missing GPU memory; you are missing a translator. The fix is a gateway in the middle that turns OpenAI-shaped calls into Anthropic-shaped ones, plus a tool that lets you point at it without recompiling anything.

The one setting, traced to the source

fazm wraps the real Claude Code agent loop in a native macOS app, and it exposes the override as a single Settings field, Custom API Endpoint, stored under the AppStorage key customApiEndpoint. On bridge start it validates the value, and if it parses, it sets one environment variable for the spawned agent process and drops the bundled key:

// Desktop/Sources/Chat/ACPBridge.swift
if let rawCustomEndpoint = defaults
    .string(forKey: "customApiEndpoint")?
    .trimmingCharacters(in: .whitespacesAndNewlines),
   !rawCustomEndpoint.isEmpty {
  if let customEndpoint = Self.validCustomAPIEndpoint(rawCustomEndpoint) {
    // every Claude-compatible call now leaves the Mac
    // pointed at your gateway, not api.anthropic.com
    env["ANTHROPIC_BASE_URL"] = customEndpoint
    // bundled key is dropped: zero built-in credits consumed
  } else {
    // malformed value is ignored, not silently mis-routed
  }
}

That is the whole mechanism. Because it is open source, you can read the exact branch in github.com/m13v/fazm at Desktop/Sources/Chat/ACPBridge.swift. Two honest constraints ship right next to the field in the app, and they are the difference between this working and silently doing nothing:

What the Settings panel tells you before you trust the override

The endpoint must speak the Anthropic API format. A raw Gemini or OpenAI key pointed straight at it will not work.
It only applies to Claude models. Pick a non-Claude model id and your request bypasses the endpoint entirely.
A malformed value (missing scheme, stray text) is ignored, not mis-routed into ANTHROPIC_BASE_URL.
When the override is on, the bundled Anthropic key is dropped, so the call never counts against built-in credits.

Putting a June 2026 trending model behind your agent

Concretely, here is the path from a name on the trending tab to a working agent turn. None of it requires touching the app source.

Confirm it is downloadable, not just trending

Open the model card. If there is no weights file or the license forbids your use, stop here. Qwen 3.7 Max-style API-only flagships do not pass this step.

Serve it with an Anthropic-compatible front

Run the weights, then put a gateway in front that exposes the Anthropic Messages API. A local LLM bridge, a corporate proxy, or a Copilot bridge all qualify, as long as the /v1/messages shape comes out the other side.

Paste the URL into Custom API Endpoint

In fazm Settings, enable Custom API Endpoint and enter your gateway URL (the field shows a host:port placeholder). The bridge restarts and writes it to ANTHROPIC_BASE_URL.

Keep a Claude model id selected and send one message

Because the override only touches Anthropic-format calls, a Claude model id must stay selected. Send a short prompt and confirm the gateway logs a request. If it sees zero traffic, you are on a non-Claude model id.

Trying to wire a trending open model into a real agent?

Bring the model and the format you are stuck on, and we will work out whether a gateway plus the custom endpoint gets you there.

Questions people actually ask about this

Frequently asked questions

Is there an official Hugging Face or GitHub list of new AI models for June 2026?

No. Neither site publishes a dated, frozen list. Hugging Face ranks models, datasets, and papers by a rolling trending score at huggingface.co/models and huggingface.co/papers/trending, and GitHub ranks repositories by a daily trending score at github.com/trending. Both feeds change every day, so any article claiming to be the definitive June list is a snapshot of one moment. The durable record of what shipped is each project's commit log and release tags, not a curated page.

What open models were showing up on the Hugging Face trending tab in June 2026?

Names that surfaced on the open-weights side included DeepSeek V4 (MIT-licensed weights on Hugging Face), the GLM series from Zhipu, Alibaba's Qwen line, Google's Gemma 4, Meta's Llama 4, and Kimi from Moonshot. Note one trap: Alibaba shipped Qwen 3.7 Max as an API-only flagship with no open weights released, so it can trend on benchmark coverage while not being something you download and run locally. Always check the model card for a weights download before assuming a trending name is self-hostable.

Why can't I just point my coding agent at a trending open model?

Because trending and runnable are different states. Most open June 2026 weights serve an OpenAI-compatible endpoint (a /v1/chat/completions path with an OpenAI messages array, the default in vLLM and SGLang). fazm's agent loop is Claude Code over ACP, which speaks the Anthropic Messages API (a /v1/messages path with Anthropic content blocks). Pointing one directly at the other returns format errors. You need a gateway in the middle that translates OpenAI to Anthropic, or a backend that natively exposes the Anthropic shape.

How does fazm actually route requests to a different model?

fazm reads a single Settings field called Custom API Endpoint, stored under the AppStorage key customApiEndpoint. On the next bridge start it validates the value and, if it parses as an http(s) URL with a host, assigns it to env["ANTHROPIC_BASE_URL"] for the spawned Claude Code process and drops the bundled Anthropic key so none of your built-in credits are spent. You can read the exact branch in Desktop/Sources/Chat/ACPBridge.swift in the public repo.

Does the custom endpoint work for any model fazm can select?

No, and the app says so in the Settings panel itself. The help text reads that the endpoint only applies to Claude models, because ANTHROPIC_BASE_URL only changes where Anthropic-format calls go. If you pick a non-Claude model id, your request bypasses the endpoint entirely and the gateway silently receives zero traffic. The fix is to keep a Claude model id selected so the agent loop routes through the override.

What is the safest way to try a June 2026 trending model with fazm?

Stand up an Anthropic-compatible gateway in front of the open model (a local LLM bridge, a corporate proxy, or a Copilot bridge all work as long as they speak the Anthropic Messages API), paste its URL into the Custom API Endpoint field, keep a Claude model id selected, and send one short message to confirm traffic reaches the gateway. Because the override is one variable, reverting is just clearing the field. Nothing about your chats, sessions, or local state changes.