Between June 1 and June 16, open-weight models crossed the line I care about most. Z.ai's GLM-5.2 shipped MIT-licensed weights that beat GPT-5.5 on SWE-bench Pro, 62.1 against 58.6, and took the top open-weights score on the Artificial Analysis Intelligence Index at 51. That is not "competitive with last year's frontier." That is ahead of a current closed flagship on the benchmark closest to real agentic coding work.
My position before the survey: the open-weight coding frontier is now real, it is mostly Chinese-led with one American exception, and the skill that mattered most this cycle was telling independent numbers from vendor-run ones, not model-picking. June produced both kinds in volume, and the community was right to be loud about the difference.
One June window
MiniMax M3 opened the month on June 1: a 428B-parameter MoE with 23B active, claiming 59.0 on SWE-bench Pro, a hair past GPT-5.5. The number was vendor-reported, no independent verification existed at launch, and coverage said so directly. The weights landed on Hugging Face about a week later, after an initial delay the community grumbled about, correctly.
NVIDIA's Nemotron 3 Ultra was unveiled at Computex the same day: 550B total parameters, 55B active, a hybrid Mamba-Transformer MoE pre-trained in NVFP4. Artificial Analysis scored it 47.7, the strongest US open-weights result, well clear of Gemma 4 31B at 39.2 and gpt-oss-120b at 33.3. It is also open in the fullest sense — weights, data, and training recipes. That matters more than it sounds like it should; I'll come back to it.
DiffusionGemma arrived June 10: Google's first open-weight diffusion LLM. Not the world's first (LLaDA-8B and Dream 7B got there earlier), but the first from a major Western lab. Apache 2.0, a 26B-class MoE that denoises 256-token blocks in parallel instead of decoding one token at a time. Google claims over 1,000 tokens per second on a single H100. That is Google's own number, so file it accordingly.
Kimi K2.7-Code shipped June 12 from Moonshot: a 1T-parameter MoE with 32B active, built around token efficiency in agentic coding, with what Moonshot claims is roughly 30% fewer thinking tokens than K2.6. Every benchmark in the launch material was Moonshot-internal, headlined by +21.8% on Moonshot's own Kimi Code Bench v2. Not one independent number existed at release.
GLM-5.2 closed the window: coding-plan subscribers got it June 13, MIT weights hit Hugging Face June 16. A 744B-parameter MoE with 1M-token context, and the scores above — with the difference that its headline placement was confirmed on an independent harness within days of the weights dropping.
Beneath the frontier noise, Qwen3.6 (the 27B dense and 35B-A3B MoE checkpoints) stayed exactly what it has been since spring: the community's default workhorse, the model that anchors the local coding roundups, and the one I actually run every day on my own Mac Studio.
And the counter-current. Meta, which seeded this entire ecosystem with roughly 1.2 billion Llama downloads, launched Muse Spark on April 8 as its first fully proprietary flagship. The company that made "open weights" a mainstream phrase exited the category in the same quarter the category reached the frontier.
The benchmark trust problem
MiniMax and Kimi both drew "wait for third-party numbers" pushback within days of launch, and both deserved it. The contrast that teaches the lesson: MiniMax's launch claim was a vendor-run 59.0 on SWE-bench Pro, edging GPT-5.5. When Artificial Analysis ran its independent composite, M3 scored 44 to GLM-5.2's 51. Those are different tests measuring different things, so it is not a caught lie — but the distance between launch-post framing and independent measurement is exactly where model strategy goes wrong.
So my rule this cycle: Artificial Analysis and independent harnesses are the reference; launch-day numbers are a press release. And the number I actually ship against is neither. It is a golden dataset in my own eval gate (I open-sourced llm-judge-evals for exactly this), because the only benchmark that decides a deployment is your tasks on your harness. GLM-5.2 earned its credibility not by claiming the top spot but by having someone else confirm it fast.
What I would build on
Qwen3.6: build on it. It runs daily on my 64GB Mac Studio under MLX and llama.cpp, and it is boring in the best sense. No drama, no surprise regressions, a deep quant ecosystem. For the local lane of a routed system, this is still the default answer.
GLM-5.2: build on it, with routing discipline. A 744B MoE is datacenter weights; in practice you consume it as a hosted API, and hosted APIs from Chinese labs carry data-residency questions that coverage flagged under China's National Intelligence Law. In the rollouts I run, that routes it away from sensitive data unless it is served from infrastructure you or a Western host controls. The MIT license is still the point. The weights are an exit hatch, and an exit hatch has value even if you never open the door.
Nemotron 3 Ultra: build on it if you sell into security review. After taking 40 enterprise GenAI deployments through security review, I can tell you the difference between "the weights are open" and "the weights, data, and training recipe are open" is a shorter meeting. A US-origin model with an independent 47.7 and full provenance is the easiest artifact in this list to get signed off.
MiniMax M3 and Kimi K2.7-Code: admire from a distance, for now. Kimi's token-efficiency premise is genuinely interesting — thinking tokens are real money in agent loops, and 30% off the bill would matter. But nothing published at release lets me verify it, and I do not put unverified models in a decision trail. When independent numbers land, I will re-evaluate. That is not cynicism; it is the same gate everything else passes.
DiffusionGemma: admire. Parallel block denoising is an architecture bet, and if the throughput claims survive independent testing, the latency math for agent loops changes. Today it is an experiment worth watching, not a workhorse.
What this changes for your model strategy
The strategy question used to be "which vendor." It is now "which weights, hosted where, evaluated on whose harness." That reframing has teeth. Closed-model pricing now negotiates against MIT-licensed weights that beat it on a public coding benchmark. Continuity risk got a live demonstration too: Anthropic's Fable 5 went dark globally for roughly eighteen days in June under a US export-control directive. Weights on disks you control do not have that failure mode. And Meta's pivot is the reminder that openness is a strategy, not a covenant — the hedge has to be weights you hold, not a vendor's stated philosophy.
What I would actually do as a team deciding this quarter: keep a closed frontier model for the hardest reasoning, stand up one open-weight lane behind a router with policy controls and a reviewable decision trail (the pattern I built Cerebellum around for local models), and gate every lane on your own evals, not anyone's launch post. You do not need to chase every June. The frontier weights will keep coming. The teams that benefit are the ones with the harness already built to measure them.