A quiet weekend on the surface hid one release that genuinely moves the board. DeepSeek promoted its V4 Flash 0731 build to production and, in the same breath, published the full weights under a permissive MIT licence, landing it among the top three open models in the world. Beyond that the weekend was about refinement rather than reinvention: xAI's Grok Imagine gained text-to-video, native 1080p and reusable character references, a real step up for a tool people already use, while OpenAI, Anthropic, Google DeepMind and the rest sat the window out after a busy week. The through-line is that the open-weights field keeps tightening, and that the most useful news right now is often a sharper version of something you already have rather than a brand new headline model.
If you build agents or run models yourself, this is the release of the weekend. DeepSeek has promoted its V4 Flash 0731 build to a production candidate on the deepseek-v4-flash API and, unusually, published the full weights under an MIT licence on the same day, which means you can use, modify and ship it commercially with no strings attached. The model keeps the same 284B total (13B active) architecture and pricing as the earlier V4 Flash, so swapping it in is a drop-in change rather than a re-integration job.
What has changed is the post-training: it scores 50 on the Artificial Analysis Intelligence Index, placing it among the top three open weights models, and it now natively supports the Responses API format and is tuned for Codex-style agent work. For anyone weighing an open alternative to closed frontier APIs, a permissively licensed model at this level of capability meaningfully shifts the maths. It surfaced first through the community, with the weights repo appearing on Hugging Face at 07:30 UTC on 31 July and being flagged by the official NVIDIA AI and Hugging Face channels before mainstream coverage caught up.
If you make short-form video, Grok Imagine has just become a good deal more useful. The 1.5 model now takes a plain text prompt straight to video, outputs native 1080p at a stated 0.25 dollars per second, and supports voice references so a character can keep the same face and voice across scenes. It also introduces multi-reference control, letting you feed up to seven visual anchors into a single generation, so you can hold a face in place while changing the location, or keep a scene while swapping the character.
Text-to-video and 1080p are generally available now on grok.com/imagine, iOS and Android, while the image and voice references started in the US for SuperGrok Heavy and Plus subscribers and are rolling out to all tiers over the following few days. For creators this narrows the gap with dedicated video tools, and doing it inside a general assistant rather than a separate app lowers the barrier to trying it. It is another in a steady run of Grok capability upgrades this week rather than a wholly new product.
Open weights are closing the gap fast. DeepSeek V4 Flash 0731 arriving under an MIT licence and immediately ranking among the top three open models means the practical case for a self-hosted alternative to closed APIs keeps getting stronger, especially for agent and coding workloads. It follows a fortnight in which Kimi K3 and Thinking Machines' Inkling-Small pushed the same way, so this is a trend with real momentum rather than a one-off.
The other pattern this weekend was doing rather than talking. The most useful news was a capability upgrade to an existing tool, Grok Imagine gaining text-to-video, 1080p and references, rather than a brand new headline model. It is also worth noting that the DeepSeek release surfaced through the community and official partner channels on X before mainstream press coverage, a reminder that watching the companies' own feeds still catches the most consequential product news first.