The open-weights wave: why every lab suddenly gives models away
Between July 14 and August 4, 2026, at least five AI labs open-sourced substantial models: MiniMax released an open-weight video model verified on consumer RTX 5090 GPUs, Higgsfield open-sourced three video models at once, Alibaba promised open weights for its 2.4T-parameter flagship Qwen 3.8-Max, Tencent's Hy-MT2 translation model passed 700K downloads, and Veed open-sourced an agentic video editor.
What exactly went open?
| Date | Company | Release |
|---|---|---|
| Jul 14 | Alibaba | Wan-Dancer-14B — open-source dance/motion video model, runs locally |
| Jul 29 | Tencent | AngelSpec — speculative-decoding framework, 1.98–2.40× speedup on Hy3-A21B |
| Jul 31 | Tencent | Hy-MT2 passes 700K downloads, #1 trending on Hugging Face |
| Aug 3 | MiniMax | Open-weight video model, verified on RTX 5090 and RTX 6000 |
| Aug 3 | Higgsfield | Hell Grind, Zephyr and Mork open-sourced, every prompt public, alongside a $1M film festival |
| Aug 3 | Alibaba | Open weights announced for Qwen 3.8-Max (2.4T) and Qwen 3.8-27B |
| Aug 4 | Veed | OpenEdit — open-source agentic video editor, installable via npm |
Why are labs giving away models that cost millions to train?
The announcements themselves telegraph the strategy. MiniMax framed its video release as "from the community, to the community" and specifically verified it runs on consumer GPUs — open weights as user-base building. Higgsfield paired its release with a $1,000,000 film festival and made every prompt public — open weights as marketing engine for its paid platform. Alibaba teased its drop through OpenRouter — open weights as distribution into the developer stack where closed rivals charge per token. In each case the weights are the top of a funnel: the paid product is the hosted API, the platform subscription, or the enterprise deal behind the free model.
Does an open video model on a gaming GPU actually matter?
Yes — it moves video generation from a metered cloud service to a local capability, the same transition image generation made when Stable Diffusion shipped in 2022. A model verified on an RTX 5090 means indie studios, VFX shops and researchers can iterate without per-second billing, fine-tune on their own footage, and ship products that don't depend on a vendor's API terms. The catch is capability lag: the open models are not the frontier ones. Seedance 2.5 and FLUX 3 Video — the models winning the platform adoption race — remain closed.
What should you watch?
Whether Alibaba actually ships Qwen 3.8-Max weights on schedule. A 2.4-trillion-parameter flagship going open would be a step-change: every previous open release at that ambition level arrived long after the closed frontier had moved on. If it lands, the open/closed capability gap collapses from years to weeks — and the pricing pressure that already pushed OpenAI to cut Luna prices 80% gets considerably worse.