Latest AI model releases, papers, and open-source projects, June 1 to 2, 2026

One model launched in this window, and it could not be downloaded. Everything that actually got an agent running on a brand-new model those two days happened at the layer the roundups never look at.

Matthew Diakonov, Written with AI

Published June 17, 20267 min read

Direct answer, verified June 17, 2026

On June 1 to 2, 2026, the one notable model launch was MiniMax M3 (June 1), and it shipped API-only, with open weights and a technical report promised on Hugging Face and GitHub about ten days later. No top lab released downloadable foundation-model weights on either day, and no single June 1 to 2 preprint reset agent work. The move that mattered those two days was at the harness layer: how you point an existing agent at a model that launched without weights. That is the spine of this page, and it traces to a single line in the Fazm source.

Primary sources: MiniMax M3 launch coverage dated June 1, 2026 (marktechpost.com, techtimes.com); the Fazm repository commit log and release tags at github.com/mediar-ai/fazm.

The dated ledger

Six entries, two days, split between the external launch and the application layer where this window actually moved. Each row is checkable against the source noted under the direct answer.

June 1Model launch
MiniMax M3 announced, API-only
Frontier-class MSA (sparse) attention model with a 1M-token context window, native multimodality, and an agentic-coding focus. Live through the company API and third-party routers on day one. Open weights and the technical report promised on Hugging Face and GitHub within about ten days, not at launch.
June 1Benchmark claim
M3 reported at 59.0% on SWE-Bench Pro
MiniMax put M3 ahead of GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%) on SWE-Bench Pro, a harder test than SWE-Bench Verified, built from 1,865 real pull requests across 41 maintained repos. All vendor-reported. No third-party Artificial Analysis or LMArena numbers at launch.
June 1Application layer
Fazm v2.9.58 fixes Claude OAuth expiring mid-query
A chat that hit an expired Claude sign-in would hang rather than re-authenticate. The fix re-emits the auth-required state during an active auth flow so the query recovers. Tagged release on github.com/mediar-ai/fazm.
June 2Paper policy (context)
No headline preprint; arXiv's hallucinated-reference ban still set the tone
No single June 1 to 2 preprint reshaped agent work. The policy story dominating arXiv was its one-year ban for submissions containing hallucinated AI references, announced in May 2026. Dated here for honesty, not as a June 1 to 2 event.
June 2Application layer
Fazm removes its bundled Anthropic API key from config
Three commits land the change: remove the bundled Anthropic key and subscription check, remove anthropic_api_key from config, remove builtin_key_blocklist. Requests now run on your own credentials or your own endpoint, the exact posture for routing to a model that launched yesterday.
June 2Application layer
Default model switched to gemini-flash-latest
Onboarding and the global shortcut now default to Gemini Flash, with non-Claude model pinning fixed so a chosen model actually sticks across a session. Visible across the v2.9.60 and v2.9.61 tags.

MiniMax M3 by the numbers, with the caveat attached

The headline number is real and the caveat is just as real. M3 led SWE-Bench Pro in MiniMax's own evaluation, on a benchmark that is genuinely harder than the saturated SWE-Bench Verified. But every figure below the context window is a vendor self-report, and on launch day no neutral lab had published a competing score.

0%SWE-Bench Pro (vendor-reported)

0MToken context window

~0Days until weights promised

0Third-party scores at launch

API-only

“On June 1, the open-weight frontier model you could not download still beat the field on a benchmark nobody else had run yet.”

MiniMax M3 launch coverage, June 1, 2026

The part every roundup skips: running it the same day

Here is the gap. A roundup tells you M3 is "open weight" and implies the path is to download it, quantize it, and load it locally. On June 1 that path did not exist, because the weights were ten days out. The path that did exist was the boring one: point your agent at the model's API. The difference between those two columns is the difference between waiting and shipping.

June 1, 2026: two reactions to an API-only launch

# June 1, 2026
# "MiniMax M3 is open weight, just download it"

$ huggingface-cli download MiniMaxAI/MiniMax-M3
Error: repo not found.
# Weights + technical report promised
# ~10 days out, not on launch day.

# Net result on June 1:
#   nothing to download
#   nothing to quantize
#   nothing to load locally
#   you wait

0% fewer lines

In Fazm the right-hand column is one Settings field. Open Settings, turn on Custom API Endpoint, and paste an Anthropic-API-compatible URL (the field placeholder is literally your-proxy:8766, with the scheme included). From that point, Claude-compatible requests route to your endpoint, your usage never touches Fazm's built-in credits, and the model id is supplied separately rather than baked into the app.

The anchor fact you can verify

The wiring lives in Desktop/Sources/Chat/ACPBridge.swift. Around line 2426 it sets env["ANTHROPIC_BASE_URL"] = customEndpoint for the agent bridge, while the model itself is passed separately as env["FAZM_SELECTED_MODEL"] read from the shortcut_selectedModel default, never a hardcoded model string. Because the app does not hardcode a model and does not hardcode an endpoint, a launch is reachable the afternoon it ships, not the version after.

One honest constraint: the custom endpoint must speak the Anthropic API format. The Settings copy says so directly, that a raw Gemini or OpenAI key will not work there. For a model like M3 that exposes a different native API, you place an Anthropic-compatible gateway (a local bridge, a corporate proxy, or a router that fronts the new model) in front of it, and Fazm talks to the gateway. That is a small integration task, and it is still same-day, which is the whole point.

Why Fazm hardened exactly this path on June 2

The timing is not a coincidence you have to take on faith, it is in the commit log. On June 2, Fazm removed its bundled Anthropic API key and the built-in key blocklist from config. The three commits are named plainly: remove the bundled Anthropic key and subscription check, remove anthropic_api_key from config, and remove builtin_key_blocklist. The effect is that requests run on your own credentials or your own endpoint by default, with nothing silently falling back to a bundled key. That is the same posture you want the instant a new model goes live: your keys, your endpoint, your call.

This is also why a two-day window is the wrong unit. The frontier feeds move every day, and most days no checkpoint you care about ships. What changes your throughput is whether your tooling can reach a new model on launch day and keep your work intact while you try it. On Fazm you fork the live chat into a new window with full prior context, point the fork at the new backend, and run the same task in both, with no auto-compacting quietly dropping the history that makes the comparison fair.

Want to run a launch-day model inside your own agent loop?

Fifteen minutes on how to wire a custom Anthropic-compatible endpoint, fork live chats across backends, and keep full context without auto-compacting.

Frequently asked questions

What AI models, papers, or open-source projects shipped on June 1 to 2, 2026?

The single notable model launch in this 48-hour window was MiniMax M3, announced June 1, 2026. It was API-only at launch: the company said the open weights and technical report would follow on Hugging Face and GitHub within roughly ten days. No top lab published new downloadable foundation-model weights dated June 1 or June 2. On the paper side, no single preprint dated to these two days reset agent work; the policy story still dominating arXiv was its one-year ban for hallucinated AI references, which had been announced in May 2026. The dated record specific to this window that you can actually read commit by commit lives at the application layer: the Fazm repo (github.com/mediar-ai/fazm) tagged v2.9.58 on June 1 and v2.9.60 and v2.9.61 across June 2, and on June 2 it removed its bundled Anthropic API key from the config.

What is MiniMax M3 and was it actually open source on June 1?

MiniMax M3 is a frontier-class model with a new MSA (sparse) attention architecture, a 1-million-token context window, native multimodality, and an agentic-coding focus. On June 1 it was reachable only through the company API and third-party routing services. The promised open weights had not been released at launch; MiniMax said the weights and a technical report would land on Hugging Face and GitHub within about ten days. So on June 1 to 2 it was an open-weight model in intent and a closed API in practice. If you read a roundup that listed M3 as a downloadable open model on June 1, that was a promise, not a file you could pull.

How good is MiniMax M3 on benchmarks, and can I trust the numbers?

MiniMax reported M3 scoring 59.0% on SWE-Bench Pro, a harder benchmark than the saturated SWE-Bench Verified, built around 1,865 real pull requests from 41 actively maintained open-source repositories. The company put GPT-5.5 at 58.6% and Gemini 3.1 Pro at 54.2% on the same test. Every one of those figures is a vendor self-report. At launch, independent third-party scores from Artificial Analysis and LMArena for M3 had not been published. Treat the 59.0% as a claim to verify once neutral evals land, not a settled result.

If a model launches API-only, how do I run it inside my existing agent the same day?

You point your agent at the model's API instead of waiting for downloadable weights. In Fazm this is one field. Settings has a Custom API Endpoint box (placeholder your-proxy:8766, scheme included). When set, Fazm injects it as the ANTHROPIC_BASE_URL environment variable for the agent bridge, so Claude-compatible requests route to your endpoint instead of the built-in path, and your usage never touches Fazm's credits. The endpoint must speak the Anthropic API format, so for a model that exposes a different native API you put an Anthropic-compatible gateway (a local bridge, a corporate proxy, or a router that fronts the new model) in front of it. The model id is never hardcoded in the app, so a launch is reachable without an app update.

Where exactly in the Fazm source does the custom endpoint get wired?

Desktop/Sources/Chat/ACPBridge.swift builds the environment for the agent bridge and sets env["ANTHROPIC_BASE_URL"] to your validated custom endpoint (around line 2426), while the selected model is passed separately as env["FAZM_SELECTED_MODEL"], read from the shortcut_selectedModel default, never a hardcoded model string. The Settings UI lives in Desktop/Sources/MainWindow/Pages/SettingsPage.swift, where the customApiEndpoint AppStorage value and the warning copy (a raw Gemini or OpenAI key will not work here, it must be Anthropic-API-compatible) are defined. It is all in a public repo, so you can read the exact behavior rather than trust a marketing line.

What did Fazm change on June 1 to 2, 2026 specifically?

On June 2, Fazm removed its bundled Anthropic API key and the built-in key blocklist from config (commits "Remove bundled Anthropic API key and subscription check," "Remove anthropic_api_key from config," and "Remove builtin_key_blocklist from config"). The practical effect: requests run on your own credentials or your own endpoint, which is the same posture you want when routing to a brand-new model. The same day it switched the default selected model to gemini-flash-latest for onboarding and shortcuts. On June 1 it shipped v2.9.58 with fixes for Claude OAuth expiring mid-query (a chat would hang instead of re-authenticating). These are tagged releases and dated commits on github.com/mediar-ai/fazm, not paraphrase.

Where do I watch the live model, paper, and open-source feeds myself?

For new and open-weight models, huggingface.co/models?sort=created. For preprints, arxiv.org/list/cs.CL/recent and arxiv.org/list/cs.LG/recent. For projects gaining traction, github.com/trending. Those three feeds move every day, which is exactly why a fixed two-day list is the wrong mental model. The durable question is not which checkpoint shipped on a given afternoon, it is whether your tooling can reach a new model the day it appears. That is the part you control.

Is Fazm open source and local, and can I verify all of this?

Yes. Fazm is a native macOS app (14.0+), fully open source on GitHub at github.com/mediar-ai/fazm, and runs locally. You bring your own Claude Pro or Max account and usage hits your existing plan. Every claim on this page about Fazm is checkable in that repo: the dated June 1 and June 2 commits, the v2.9.58 to v2.9.61 tags, and the ANTHROPIC_BASE_URL wiring in ACPBridge.swift. The external claims trace to the launch coverage of MiniMax M3 dated June 1, 2026.

More on the harness layer, dated windows, and reaching new models the day they ship.