Who Is Still Shipping Open Weights in August 2026

If you run models on infrastructure you control, the question that matters is not which model tops a leaderboard. It is who will still be publishing weights in eighteen months, because that determines whether your setup gets updates or slowly becomes an unmaintained dependency.

The answer changed materially during 2026. Here is the state of it as of early August, with the caveat that this is a snapshot of a fast-moving supply and should be re-checked rather than trusted a year from now. The live version of this list is the model hub on Hugging Face, where the weights and license files actually appear.

Actively Publishing

Alibaba (Qwen). The most prolific open publisher by volume and, by download counts, the most used family in the world. Ships across the full size range from under a billion parameters to hundreds of billions, in dense and mixture-of-experts variants, with separate instruct, thinking, coding, and multimodal lines. Announced in August 2026 that its flagship Qwen3.8-Max would be the first Max-tier model released openly. Mostly Apache 2.0. The navigation guide is the Qwen open-weights family explained.

DeepSeek. Ships frontier-adjacent mixture-of-experts models under MIT, most recently the V4 Flash line in July 2026. Large, server-scale, permissively licensed.

Google (Gemma). Gemma 4 landed in April 2026 across four sizes from phone-scale to 31 billion parameters, and moved the family to Apache 2.0 for the first time. The small end is directly relevant to local writing.

Moonshot (Kimi). Publishes very large models, including a 2.8 trillion parameter release in July 2026 with a 1 million token context. Impressive, and beyond any hardware you own.

Tencent, and various Chinese labs. Continue to publish across sizes, with licenses that vary and are worth reading individually.

Mistral. The main European publisher. Confirmed a new open-weight model entering early access in July 2026 as part of a new family, with a broader release expected later in the year. Parameter counts, benchmarks, and license terms were not disclosed at early access, so treat details as unsettled. The sovereignty argument around it is in Mistral open models and European data sovereignty.

OpenAI. Maintains the gpt-oss line, a deliberate and limited open contribution alongside a closed flagship business. Covered in gpt-oss: OpenAI’s open weights explained.

Smaller labs. Thinking Machines published a large mixture-of-experts model in July 2026, one of the few US-built open entries at that scale. NVIDIA continues to publish small models aimed at deployment.

No Longer Publishing Frontier Weights

Meta. Has not shipped a new open flagship Llama in over a year, and the frontier work now ships as a closed API line. Existing Llama weights and licenses are unaffected, so nothing you run breaks. What you lose is a successor. The full picture is in Meta went closed.

Anthropic. Has never published open weights and shows no sign of starting.

What The Shape Of This Tells You

Three things, stated as plainly as the evidence supports.

Supply is healthy by volume. More good open models shipped in the first seven months of 2026 than in all of 2025. The ecosystem is not in decline.

Supply is geographically concentrated. The majority of the significant open releases in this period came from Chinese labs. That is a procurement fact, not a quality judgment, and it is worth naming in a policy document rather than discovering during a review. Running weights locally means no data reaches the developer, so the honest question is licensing, provenance, and organizational policy rather than data transfer. Chinese AI models: what professionals should know works through it.

The small end is safer than the frontier. Multiple families ship well-maintained small models under permissive licenses. If your use case is writing, you are choosing from an unusually robust part of the market, and you have real alternatives if any single publisher stops.

What This Means For Choosing

For local writing, three criteria in order:

1. Permissive license. Apache 2.0 or MIT means no negotiation and no field-of-use surprises.
2. Size that fits your actual machine. For rewriting, 1 to 4 billion parameters quantized. A model you cannot load is not a model you have. How much RAM to run a local LLM gives the arithmetic.
3. An active line, not a single release. A family with three versions in eighteen months is more likely to ship a fourth.

Notice what is not on the list: benchmark rank, parameter count, and which country the lab is in. For a constrained transformation like rewriting, the first two barely move the result and the third is a policy question rather than a technical one.

The Argument That Does Not Depend On The Scoreboard

Every entry above could change next quarter. A lab that publishes today can stop, as Meta did. A lab that has never published can start.

That volatility is the argument for open weights rather than against it. Weights you hold cannot be withdrawn, repriced, retired, or rerouted. Everybody who depended on a hosted Llama API had to migrate when Meta changed direction. Everybody who had the file on disk did not.

A Wrivio Context for a model policy note could say:

Rewrite this as a factual internal policy note. Neutral register, complete sentences. State the decision, the criteria, and the review date. Keep every model name, license name, and date exactly as written. Do not add recommendations, risk language, or claims about vendors that are not in the original.

Press Ctrl+Shift+Space, paste the draft, and check the diff. Watch specifically for the rewrite generalizing a license name, because “open source” and “open weights” are not interchangeable and a policy document that confuses them creates real problems later.

Common Questions

Is the open weights ecosystem shrinking?

No. Release volume and quality in 2026 exceed 2025. What changed is which labs are producing them, with Meta exiting the frontier open tier and Chinese labs supplying most of it.

Does it matter which country published the model I run locally?

For data flow, no: weights running on your machine send nothing anywhere. For procurement policy, licensing, and provenance requirements, it can matter, so check your own organization’s rules.

Which open models are actually usable for writing on a laptop?

Small dense models in the 1 to 4 billion parameter range, quantized to four bits. Current Qwen and Gemma small models are the mainstream choices.

How often should I re-check this?

Twice a year is enough for a writing setup. The frontier moves weekly and the small end moves slowly, which is one more reason the small end is a comfortable place to be.

Download Wrivio for Windows to run a permissively licensed small model locally, on weights that stay on your disk.