Hello, dear readers. Happy belated Thanksgiving and Black Friday!
This year, it felt like I was living in a permanent DevDay. Every week, some lab releases a new model, a new agent framework, or a new “this changes everything” demo. It’s overwhelming. But this is also the first year that I feel like AI is finally diversifying. Instead of one or two frontier models on the cloud, the entire ecosystem is becoming more diverse: open and closed, big and small, Western and Chinese, cloud and local.
So, in this Thanksgiving edition, I’ll share with you what I’m really thankful for in AI in 2025: releases that I think will be important not just during this week’s hype cycle, but in the next 12 to 24 months.
1. OpenAI continues to ship well: GPT-5, GPT-5.1, Atlas, Sora 2, and Open Weight
As a company that has undoubtedly created "Generation AI" OpenAI, which launched its blockbuster product ChatGPT in late 2022, had perhaps one of the toughest challenges of any AI company in 2025. It’s about continuing its growth trajectory even as well-funded competitors like Google with its Gemini model and other startups like Anthropic roll out competitive products of their own.
Thankfully, OpenAI has risen to this challenge as well. The highlight was GPT-5, announced in August as the next frontier inference model. Then in November, we announced GPT-5.1 with new Instant and Thinking variants that dynamically adjust the “think time” spent on each task.
In fact, the release of GPT-5 was eventful. VentureBeat chronicled early math and coding failures and that the community’s reaction was more subdued than expected." However, it was quickly fixed based on user feedback and I am personally happy and impressed with it as I use this model on a daily basis.
At the same time, companies that actually use this model report solid profits. For example, according to ZenDesk Global, agents using GPT-5 now resolve more than half of their customer tickets, with some customers seeing resolution rates of 80-90%. This is a quiet story. These models don’t necessarily make a good impression on the chatty class of X, but they start to drive real KPIs.
On the tools side, OpenAI finally provides developers with a full-fledged AI engineer with GPT-5.1-Codex-Max. GPT-5.1-Codex-Max is a new coding model that can run long-running agent workflows and is already the default in OpenAI’s Codex environment. VentureBeat covers it in detail in “OpenAI Debuts GPT-5.1-Codex-Max Coding Model, Already Completed 24 Hours of Tasks Internally.”
Additionally, there is ChatGPT Atlas, a full browser with ChatGPT built into Chrome itself. Sidebar overview, on-page analysis, and search are tightly integrated into normal browsing. This is the clearest sign yet that Assistant and Browser are colliding.
On the media side, Sora 2 turns the original Sora video demo into a complete video and audio model with better physics, synchronized sound and dialogue, and more control over style and shot structure. Additionally, the dedicated Sora app with a full-fledged social networking component now allows any user to create their own TV network in their pocket.
Finally, and perhaps most emblematically, OpenAI released gpt-oss-120B and gpt-oss-20B, open-weight MoE inference models under Apache 2.0-style licenses. No matter what you think of its quality (and early open source users complained loudly), this is the first time since GPT-2 that OpenAI has really focused on the public commons.
2. China’s open source wave goes mainstream
If 2023-2024 is about Rama and Mistral, 2025 will be about China’s promiscuous ecosystem.
A study by MIT and Hugging Face found that China has a slight lead over the US in global open model downloads, thanks in large part to DeepSeek and Alibaba’s Qwen family.
Highlights:
-
Deep Seek-R1 It was released in January as an open source inference model comparable to OpenAI’s o1. Featuring the weight of an MIT license and a family of refined small models. VentureBeat tracked the story from release to cybersecurity implications for the performance-tuned R1 variant.
-
Kimi K2 thinking Moonshot is a “thinking” open source model that uses tools to reason step-by-step, very similar to the o1/R1 mold, and has been positioned as the best open reasoning model in the world to date.
-
Zai We shipped GLM-4.5 and GLM-4.5-Air as “agent” models and open sourced the base and hybrid inference variants on GitHub.
-
Baidu’s Ernie 4.5 The family debuted as a fully open source multimodal MoE suite under Apache 2.0. It includes a 0.3B dense model and a visual “thinking” variant that focuses on charts, STEM, and tool usage.
-
alibaba’s Quen 3 This line, which includes Qwen3-Coder, large-scale inference models, and the Qwen3-VL series released in the summer and fall of 2025, continues to set a high bar for open weight in coding, translation, and multimodal inference, and this summer I declared: "
Kwen’s summer."
VentureBeat is tracking these changes, including Chinese math and reasoning models like the Light-R1-32B and Weibo’s smaller VibeThinker-1.5B. These models outperform DeepSeek’s baseline with a small training budget.
If you’re interested in open ecosystems and on-premises options, this is the year that China’s open ended scene became more than just a curiosity and became a full-fledged alternative.
3. Small, local models will grow
Another thing I’m grateful for is that finally good Not just toys, but also small models.
In 2025, Liquid AI has been working on advancing Liquid Foundation Models (LFM2) and the LFM2-VL vision language variant. They were designed from day one for low-latency, device-enabled deployments, including large clusters as well as edge boxes, robots, and constrained servers. The new LFM2-VL-3B targets embedded robotics and industrial autonomy and is scheduled for demonstration at ROSCon.
On the big tech side, Google’s Gemma 3 series made a strong case that you can be “small” and still do the job. Gemma 3 ranges in parameters from 270M to 27B, with all larger variations featuring open weights and multimodal support.
A standout is the Gemma 3 270M. It’s a compact model purpose-built for fine-tuning and structured text tasks (think custom formatters, routers, and watchdogs), and has been featured both on Google’s developer blog and in community discussions in local LLM circles.
These models may never be a trend in X, but they are exactly what you need for privacy-sensitive workloads, offline workflows, thin client devices, and “agent fleets” where you don’t want every tool call to hit a giant frontier LLM.
4. Meta + Mid-Journey: Aesthetics as a Service
One of the strange developments this year is that the meta has partnered with Midjourney rather than just trying to take it down.
In August, Meta announced a deal to license Midjourney’s “aesthetic technology,” or image and video generation stack, and integrate it into Meta’s future models and products, from Facebook and Instagram feeds to Meta AI capabilities.
VentureBeat covered the partnership in “Meta partners with Midjourney to license its technology for future models and products,” raising the obvious question. Will this delay or restructure Midjourney’s own API roadmap? We’re still waiting for an answer, but unfortunately the stated plans for API releases have yet to come to fruition, suggesting that they are.
But for creators and brands, the immediate implications are simple. Midjourney-grade visuals will start appearing on mainstream social tools instead of being locked away in Discord bots. That could normalize high-quality AI art for a wider audience, forcing rivals like OpenAI, Google, and Black Forest Labs to continue raising the bar.
5. Google’s Gemini 3 and Nano Banana Pro
Google tried to answer GPT-5 with Gemini 3. Gemini 3 is touted as the most capable model to date, with better reasoning, coding, and multimodal understanding, as well as a new Deep Think mode for slow and difficult problems.
VentureBeat’s coverage, “Google unveils Gemini 3 to claim lead in math, science, multimodal, agent AI,” is framed as a direct focus on frontier benchmarks and agent workflows.
But the surprise hit is Google’s new flagship image generator, Nano Banana Pro (Gemini 3 Pro Image). We specialize in infographics, diagrams, multi-subject scenes, and multilingual text that is rendered legible in 2K and 4K resolutions.
In the world of enterprise AI, charts, product schematics, and images that “visually explain the system” are more important than fantasy dragons. This is a big problem.
6. Wildcards to watch
I’d appreciate a few more releases, even if they don’t fit neatly into one bucket.
-
Flux.2 by Black Forest Labs The Image model was launched earlier this week with the ambition to challenge both Nano Banana Pro and Midjourney in terms of quality and management. VentureBeat dug into the details in “Black Forest Labs unveils Flux.2 AI image model to take on Nano Banana Pro and Midjourney.”"
-
Anthropic’s Claude Works 4.5is a new flagship aimed at cheaper, more competent coding and long-term task execution, featured in “Introducing Anthropic’s Claude Opus 4.5: Cheaper AI, Infinite Chat, and Better-than-Human Coding Skills.”"
-
The steady drumbeat of open math/inference models, from Light-R1 to VibeThinker and more, shows that it doesn’t take a $100 million training run to move the needle.
Last thoughts (so far)
If 2024 was the year of “one big model in the cloud,” 2025 is the year the map exploded. Multiple frontiers will culminate, China will lead with an open model, small and efficient systems will rapidly mature, and creative ecosystems like Midjourney will be drawn into the big tech stack.
I appreciate the fact that we now have more than just a single model. option — closed and open, local and hosted, inference-first and media-first. For journalists, builders and businesses, that diversity is the real story of 2025.
I hope you have a nice vacation. Best vacation to you and your loved ones!
