Editorial · 9 min · May 12, 2026

The state of open weights in May 2026

Five frontier-class open-weight releases shipped in the last thirty days. Three of them came from Chinese labs. A short tour of where the field actually is, written for readers who already know that “open weights” and “open source” do not mean the same thing.

The headline release schedule, with no decoration

In April and the first half of May 2026, the open weights calendar looked like this. Meta released Llama 4 with the Scout and Maverick variants, the first pushing context windows to a claimed 10 million tokens through a mixture-of-experts architecture. Alibaba released Qwen 3.5 across the usual ladder of sizes, with the smaller variants keeping Apache 2.0 licensing. DeepSeek shipped V4 in Pro and Flash tiers, and the smaller MIT-licensed core continues to be the most permissive frontier-grade model anyone has released. Google followed with Gemma 4, leaning on on-device deployment and tight Apple Silicon performance. Mistral, the only European lab in the top tier, pushed out Mistral Medium 3.5 on April 29 with Apache 2.0 weights and SWE-Bench Verified numbers that put it credibly in the conversation.

Five releases in thirty days is not a coincidence. The pattern, visible for at least two release cycles now, is that the open-weight calendar has effectively become quarterly. Labs ship in clusters because the cost of being last in a thirty-day window is much smaller than the cost of being last in a six-month window.

The benchmark numbers, and what they actually mean

Benchmark season produced the expected wave of claims. DeepSeek V4 Pro reached 80.6 on SWE-Bench Verified and 90.1 on GPQA Diamond, with a 1 million token context. Mistral Medium 3.5 came in at 77.6 on SWE-Bench, the strongest non-Chinese, non-American open-weight result of the cycle. Qwen 3.5 and Llama 4 trade leadership depending on which evaluation suite you look at, and Gemma 4 sits behind the frontier on pure capability while quietly winning the laptop deployment story.

Numbers from labs are worth what a vendor benchmark is usually worth. Independent leaderboards have caught up faster than they used to, and the ranking is more stable than a year ago, but the gap between leaderboard performance and what a model actually does on your workload remains. Running the same prompt set against three candidates in LM Studio tells you more in an afternoon than a week of reading evaluation papers.

The licensing layer, where the actual freedom lives

Capability is one axis, license is another, and the two do not correlate the way the marketing suggests. If you need an OSI-approved license for legal reasons, the shortlist gets very short very fast: DeepSeek V4 (MIT) for the frontier-tier slot, smaller Qwen sizes (Apache 2.0) for mid-tier coverage, OLMo from the Allen Institute for full transparency, and EuroLLM-22B for the rare combination of Apache 2.0 weights with a fully published training pipeline.

The middle tier of licenses (Llama Community License, Gemma Terms of Use, the various bespoke documents Chinese labs publish) cover most commercial settings comfortably, but they all carry restrictions that a paranoid legal team will flag: MAU caps, training-restriction clauses, export concerns, sometimes ambiguity about whose jurisdiction governs. The work of reading those terms is unglamorous and unavoidable.

The geographic shape of the field

Three observations that the 2026 release cycle made unavoidable. Chinese labs are no longer the budget-tier alternative. DeepSeek, Qwen, Kimi and GLM hold leaderboard positions that the American labs cannot consistently beat with open weights. The Mistral exception is real and important, but it does not change the basic shape of the field. Calls for “European AI sovereignty” that lean on the assumption that the open weights everyone runs are European need to do the math again.

American labs continue to lead on closed models, on infrastructure, and on the boring but essential layer of tooling around weights: evaluation, safety, deployment. The open weights they ship are increasingly a second-tier product strategically, not a first-tier contribution. Meta is the partial exception. Google ships open weights in a way that feels designed to be useful on Pixel and Chromebook hardware specifically.

Europe has one credible frontier-tier lab and a research consortium producing genuinely open models like EuroLLM. That is not nothing. Nor is it the kind of footprint that justifies the policy rhetoric. The honest reading is that European AI sovereignty, to the extent it is achievable, will be a sovereignty of deployment, governance, and data residency rather than a sovereignty of the underlying weights.

What is worth running today

Three practical recommendations, calibrated to actual hardware most readers have available.

For a 16 GB laptop, Qwen 3.5 7B in Q4_K_M is the best general-purpose choice. It handles chat, summarization, light coding, and most retrieval workloads. Phi-5 7B is the credible runner-up if you need slightly faster inference at a small quality cost.

For a 24 to 32 GB workstation, Qwen 3.5 14B or Llama 4 Scout at moderate context lengths give you a real working assistant for substantial coding and analysis. The DeepSeek V4 distilled variants are worth a look if you do math-heavy work.

For a multi-GPU server with serious VRAM (96 GB or more), DeepSeek V4 Pro is the frontier-tier open choice today. Llama 4 Maverick competes on most tasks, and Mistral Medium 3.5 is the one to test if your legal team prefers EU-origin weights.

What to watch over the next quarter

Three specific things, none of them speculative. The next OpenEuroLLM release, which is supposed to ship on the Jupiter exascale system in Germany and which will be the first time a European public-money model competes at scale. The first 1M-context Apache 2.0 release from a Western lab, if it happens. And the rate at which Chinese labs continue to release under MIT or Apache, because the political pressure to tighten those licenses is real and the labs have so far ignored it.

The field is healthier than it has ever been, which is not the same as saying it is settled. Three years from now, the leaderboard will look different and at least one of the labs in the table above will have slipped or stopped. The good news is that open weights are not a subscription. Whatever you download today will keep running.