For most of the last three years, the answer to what open weight model should we build on was Llama. That answer has quietly stopped being available.

Meta Superintelligence Labs introduced Muse Spark on 8 April 2026 and has since added Muse Image and Muse Code. The Muse family replaced Llama. Llama 5 was never released. llama.com now redirects to Meta’s developer site. And Muse Spark is proprietary, distributed free through the Meta AI app, WhatsApp, Instagram, Facebook, and Meta AI glasses, with API access limited to a private preview and only a stated hope to open source future versions.

Meta did not announce a strategy reversal. It simply shipped a different kind of product and let the old one lapse.

What Meta appears to be optimising for now

Muse Spark was announced with no benchmark scores, no disclosed context window, and no published API pricing. Meta describes it as small and fast by design, able to see and understand what the user is looking at.

Read that alongside the distribution and the intent is legible. This is not a model competing for developer mindshare on Hugging Face. It is an assistant embedded in apps that several billion people already have installed, on hardware Meta also sells. Open weights served a purpose when Meta was behind and needed an ecosystem. Universal free distribution serves a different purpose, and the two strategies do not obviously need each other.

Where the open weight frontier actually is

The gap closed from the other direction, quickly.

DeepSeek-V4-Flash, released into public beta on 31 July 2026, is a 304 billion parameter mixture of experts model with a million token context under plain MIT. DeepSeek’s own card reports 82.7 on Terminal Bench 2.1 and states it outperforms the larger V4-Pro preview.

Kimi K3 from Moonshot AI is a 2.8 trillion parameter model with a million token context, described as the first open model at 3 trillion class scale, with a reported 93.5 percent on GPQA Diamond. It ships under a bespoke Kimi K3 Licence rather than a standard one.

Qwen3.8-Max from Alibaba, announced 3 August 2026, reports the strongest agentic coding figures currently published anywhere, at 86.6 on Terminal-Bench 2.1. Alibaba committed to releasing open weights within a week of the announcement. At the time of writing we could not confirm they have shipped, so we do not yet describe it as open.

Mistral Medium 3.5 is the outlier and arguably the most practical. A dense 128 billion parameter model with a 256,000 token context, released under a modified MIT licence, designed to run on as few as four GPUs. It replaced Devstral 2 as the default in Mistral’s own coding agents.

The licence problem this creates

Llama’s licence had known constraints, but it was one licence that a great many legal teams had already reviewed. That shared foundation is gone. In its place are plain MIT from DeepSeek, a bespoke licence from Moonshot, a modified MIT from Mistral, and an unconfirmed commitment from Alibaba.

For a team that treated open weights as a single procurement decision, this is now four separate decisions. The practical effect is that DeepSeek’s plain MIT is worth more than its benchmark position alone would suggest, simply because it is the one that does not need a legal review.

What to do if you built on Llama

Nothing is broken. Weights you already hold under the Llama licence remain usable under the terms you accepted. What has changed is the roadmap: there will be no successor from Meta, so any plan that assumed one needs a new foundation.

If your priority is licence simplicity, start with DeepSeek-V4-Flash. If it is running on hardware you can actually afford, start with Mistral Medium 3.5. If it is raw capability and you have the legal capacity to read a bespoke licence, look at Kimi K3. And watch whether Alibaba honours the Qwen3.8-Max open weights commitment, because if it does, the strongest reported agentic coding model in the world becomes something you can run yourself.