Hello, dear readers. Elated belated Thanksgiving and Black Friday!
This year I felt like I was living in a indefinite Deva Day. Every week, some lab releases a modern model, a modern agent platform, or a modern “this changes everything” demo. It’s overwhelming. But this is also the first year that I feel like AI is finally diversifying — not just one or two pioneering cloud models, but the entire ecosystem: open and closed, giant and petite, Western and Chinese, cloud and on-premises.
Here’s what I’m really grateful for about the 2025 AI Thanksgiving release in 2025 – releases that feel like they’re going to make a difference over the next 12-24 months, not just during this week’s hype cycle.
1. OpenAI maintained good shipping: GPT-5, GPT-5.1, Atlas, Sora 2 and open scales
As the company that undeniably ushered in the era of “generative AI” with its viral ChatGPT product in delayed 2022, OpenAI likely had one of the most hard tasks for any AI company in 2025: continuing its growth trajectory even as well-funded competitors like Google with its Gemini models and other startups like Anthropic introduced their own highly competitive offers.
Fortunately, OpenAI rose to the challenge, and quite a bit. Its main act was GPT-5, unveiled in August as another pioneering reasoning model, followed in November by GPT-5.1 with modern Instant and Thinking variants that dynamically adjust the amount of “thinking time” spent on a task.
In practice, the GPT-5 launch has been bumpy – VentureBeat documented early math and coding errors and better-than-expected community response, concluding that “OpenAI’s GPT-5 rollout has not been smooth,” but this was quickly corrected based on user feedback, and as an everyday user of the model, I’m personally pleased and impressed with it.
At the same time, companies that actually operate these models are seeing solid profits. Global ZenDeskfor example, it claims that agents using GPT-5 now resolve more than half of customer tickets, with some customers seeing resolution rates of 80-90%. It’s a peaceful story: these models don’t always impress the chattering classes at X, but they’re starting to move real KPIs.
On the tooling side, OpenAI has finally given developers a stern AI engineer with GPT-5.1-Codex-Max, a modern coding model that can execute long, agent-based workflows and is already OpenAI’s default Codex framework. VentureBeat detailed this in “OpenAI debuts GPT-5.1-Codex-Max encoding model and has already completed the 24-hour task internally.”
Then there’s ChatGPT Atlas, a full browser with ChatGPT built right into Chrome itself – sidebar summaries, on-page analysis, and search tightly integrated into regular browsing. This is the clearest sign that “assistant” and “browser” are on a collision course.
On the media side, Sora 2 transformed the original Sora video demo into a full video and audio model with better physics, synchronized audio and dialogue, and more control over the style and structure of shots, as well as a dedicated Sora app with a full-fledged social networking component, allowing any user create your own TV network in your pocket.
Finally – and perhaps most symbolically – OpenAI released gpt-oss-120B and gpt-oss-20B, open-weight MoE inference models under an Apache 2.0-style license. Regardless of what you think of their quality (and early adopters of open source have been vocal in their complaints), this is the first time since GPT-2 that OpenAI has given stern attention to the public space.
2. China’s open source wave hits the mainstream
If 2023-24 was about Llama and Mistral, 2025 belongs to China’s openweight ecosystem.
This was demonstrated by a study conducted by MIT and Hugging Face China currently slightly leads the United States in global downloads under the open modelthanks in enormous part to DeepSeek and Alibaba’s Qwen family.
Overview of the most significant events:
-
DeepSeek-R1 launched in January as an open-source inference model competing with OpenAI’s o1, with MIT-licensed weights and a family of smaller models. VentureBeat followed the story from launch, through its impact on cybersecurity, to performance-optimized R1 variants.
-
Kimi K2’s thinking from Moonshot, an open source “thinking” model that reasons step by step with tools, very much in the form of o1/R1, and is positioned as the best open reasoning model in the world to date.
-
Z.ai delivered GLM-4.5 and GLM-4.5-Air as “agent” models, open source databases, and hybrid reasoning variants on GitHub.
-
Baidu ERNIE 4.5 Family emerged as a fully open, multimodal MoE package under Apache 2.0, featuring a dense 0.3B model and visual “Thinking” variants focusing on graphs, STEM, and tool operate.
-
Alibaba Qwen3 line – which includes Qwen3-Coder, enormous inference models, and the Qwen3-VL series released in summer and fall 2025 – continues to set a high bar for open scales in coding, translation, and multimodal reasoning, which led me to consider this past summer as “
Summer Qwen.”
VentureBeat is tracking these developments, including Chinese math and reasoning models like Delicate-R1-32B and Weibo’s miniature VibeThinker-1.5B, which are outperforming DeepSeek’s baselines on modest training budgets.
If you care about open ecosystems or on-premise options, this is the year when the Chinese open-weight scene stopped being a curiosity and became a stern alternative.
3. Miniature and local models grow up
Another thing I’m grateful for: we finally get it Good petite models, not just toys.
Liquid AI spent 2025 developing variants of the Liquid Foundation Models (LFM2) and LFM2-VL vision language, designed from day one for low-latency and device-aware deployments – edge devices, robots and constrained servers, not just giant clusters. Newer LFM2-VL-3B focuses on embedded robotics and industrial autonomy, with demonstrations scheduled for ROSCon.
When it comes to gigantic tech, Google’s Gemma 3 line is sturdy proof that “small” can still have possibilities. Gemma 3 covers parameters from 270M to 27B, all with open scales and multimodal support on larger variants.
A standout is the Gemma 3 270M, a compact model built specifically for tuning and structured text tasks – such as custom formatters, routers, and watchdogs – discussed both on the Google Developer Blog and in community discussions in local LLM circles.
These models may never gain traction in X, but they’re exactly what you need for privacy-sensitive workloads, offline workflows, thin-client devices, and “agent swarms” where you don’t want every tool call going to a giant borderline LLM.
4. Meta + Midjourney: aesthetics as a service
One of the weirder twists this year: Meta teamed up with Midjourney instead of just trying to beat him.
In August, Meta announced a deal to license Midjourney’s “aesthetic technology” – an image and video generation stack – and to integrate it with future Meta models and products, from Facebook and Instagram feeds to Meta AI features.
VentureBeat covered the partnership in the article “Meta partners with Midjourney and will license its technology for future models and products,” raising an obvious question: Does this tardy down or change Midjourney’s API roadmap? I’m still waiting for a response, but unfortunately it states that API release plans have yet to materialize, which suggests that it has.
But for creators and brands, the immediate takeaway is basic: Midjourney-level visuals are starting to appear across mainstream social tools, rather than being locked away in the Discord bot. This could normalize higher-quality AI art for a much wider audience and force rivals like OpenAI, Google and Black Forest Labs to continually raise the bar.
5. Google Gemini 3 and Nano Banana Pro
Google tried to respond to GPT-5 with Gemini 3, touted as the most powerful model ever, with better reasoning, coding and understanding of multimodality, as well as a modern Deep Think mode for tardy and hard problems.
VentureBeat’s coverage of “Google unveils Gemini 3 as it emerges as a leader in math, science, multimodal and agent-based AI” framed it as a direct shot at pioneering benchmarks and agent-based workflows.
However, the surprise hit is Nano Banana Pro (Gemini 3 Pro Image), Google’s modern flagship image generator. It specializes in infographics, diagrams, multi-topic scenes and multilingual text that is actually readable in 2K and 4K resolutions.
In the world of enterprise AI — where charts, product diagrams, and images that “visually explain the system” matter more than fantasy dragons — that’s a gigantic deal.
6. Wild cards I pay attention to
A few more releases I’m grateful for, even if they don’t fit neatly into one bucket:
-
Flux.2 by Black Forest Labs image models that launched just earlier this week with the goal of challenging both Nano Banana Pro and Midjourney in terms of quality and control. VentureBeat delved into the details in the article “Black Forest Labs Launches Flux.2 AI Image Models to Challenge Nano Banana Pro and Midjourney.”
-
Claude Opus by Anthropic 4.5a modern flagship product aimed at cheaper, more productive coding and long-term task execution, described in the article “Anthropic’s Claude Opus 4.5 is here: Cheaper AI, infinite chats and coding skills that beat humans.”
-
The constant heartbeat of open math/inference models – from Delicate-R1 to VibeThinker and others – that show you don’t need $100 million worth of training to move the needle.
Final thought (for now)
If 2024 was the year of “one big cloud model,” then 2025 is the year the map exploded: multiple frontiers at the top, China leading the way in open models, petite and productive systems maturing rapidly, and imaginative ecosystems like Midjourney being pulled into the gigantic tech stacks.
I’m grateful not just for any single model, but that we now have it options — closed and open, local and hosted, reasoning first and media first. For journalists, designers and companies, this diversity is the real story of 2025.
Merry Christmas and all the best to you and your loved ones!