Hi there, pricey readers. Joyful belated Thanksgiving and Black Friday!
This 12 months has felt like residing inside a everlasting DevDay. Each week, some lab drops a brand new mannequin, a brand new agent framework, or a brand new “this adjustments all the things” demo. It’s overwhelming. Nevertheless it’s additionally the primary 12 months I’ve felt like AI is lastly diversifying — not only one or two frontier fashions within the cloud, however an entire ecosystem: open and closed, large and tiny, Western and Chinese language, cloud and native.
So for this Thanksgiving version, right here’s what I’m genuinely grateful for in AI in 2025 — the releases that really feel like they’ll matter in 12–24 months, not simply throughout this week’s hype cycle.
1. OpenAI stored transport robust: GPT-5, GPT-5.1, Atlas, Sora 2 and open weights
As the corporate that undeniably birthed the “generative AI” period with its viral hit product ChatGPT in late 2022, OpenAI arguably had among the many hardest duties of any AI firm in 2025: proceed its progress trajectory whilst well-funded rivals like Google with its Gemini fashions and different startups like Anthropic fielded their very own extremely aggressive choices.
Fortunately, OpenAI rose to the problem after which some. Its headline act was GPT-5, unveiled in August as the subsequent frontier reasoning mannequin, adopted in November by GPT-5.1 with new Instantaneous and Considering variants that dynamically alter how a lot “considering time” they spend per activity.
In follow, GPT-5’s launch was bumpy — VentureBeat documented early math and coding failures and a cooler-than-expected neighborhood response in “OpenAI’s GPT-5 rollout will not be going easily,” nevertheless it rapidly course corrected based mostly on consumer suggestions and, as a day by day consumer of this mannequin, I am personally happy with it and impressed with it.
On the identical time, enterprises truly utilizing the fashions are reporting strong good points. ZenDesk Global, for instance, says GPT-5-powered brokers now resolve greater than half of buyer tickets, with some clients seeing 80–90% decision charges. That’s the quiet story: these fashions could not at all times impress the chattering courses on X, however they’re beginning to transfer actual KPIs.
On the tooling aspect, OpenAI lastly gave builders a critical AI engineer with GPT-5.1-Codex-Max, a brand new coding mannequin that may run lengthy, agentic workflows and is already the default in OpenAI’s Codex surroundings. VentureBeat lined it intimately in “OpenAI debuts GPT-5.1-Codex-Max coding mannequin and it already accomplished a 24-hour activity internally.”
Then there’s ChatGPT Atlas, a full browser with ChatGPT baked into the chrome itself — sidebar summaries, on-page evaluation, and search tightly built-in into common looking. It’s the clearest signal but that “assistant” and “browser” are on a collision course.
On the media aspect, Sora 2 turned the unique Sora video demo right into a full video-and-audio mannequin with higher physics, synchronized sound and dialogue, and extra management over type and shot construction, plus a devoted Sora app with a full fledged social networking part, permitting any consumer to create their own TV network in their pocket.
Lastly — and perhaps most symbolically — OpenAI launched gpt-oss-120B and gpt-oss-20B, open-weight MoE reasoning fashions below an Apache 2.0–type license. No matter you consider their high quality (and early open-source customers have been loud about their complaints), that is the primary time since GPT-2 that OpenAI has put critical weights into the general public commons.
2. China’s open-source wave goes mainstream
If 2023–24 was about Llama and Mistral, 2025 belongs to China’s open-weight ecosystem.
A examine from MIT and Hugging Face discovered that China now slightly leads the U.S. in global open-model downloads, largely due to DeepSeek and Alibaba’s Qwen household.
Highlights:
-
DeepSeek-R1 dropped in January as an open-source reasoning mannequin rivaling OpenAI’s o1, with MIT-licensed weights and a household of distilled smaller fashions. VentureBeat has adopted the story from its launch to its cybersecurity influence to performance-tuned R1 variants.
-
Kimi K2 Considering from Moonshot, a “considering” open-source mannequin that causes step-by-step with instruments, very a lot within the o1/R1 mould, and is positioned as the perfect open reasoning mannequin to this point on the earth.
-
Z.ai shipped GLM-4.5 and GLM-4.5-Air as “agentic” fashions, open-sourcing base and hybrid reasoning variants on GitHub.
-
Baidu’s ERNIE 4.5 household arrived as a totally open-sourced, multimodal MoE suite below Apache 2.0, together with a 0.3B dense mannequin and visible “Considering” variants centered on charts, STEM, and power use.
-
Alibaba’s Qwen3 line — together with Qwen3-Coder, massive reasoning fashions, and the Qwen3-VL sequence launched over the summer season and fall months of 2025 — continues to set a excessive bar for open weights in coding, translation, and multimodal reasoning, main me to declare this previous summer season as “
Qwen’s summer season.”
VentureBeat has been monitoring these shifts, together with Chinese language math and reasoning fashions like Mild-R1-32B and Weibo’s tiny VibeThinker-1.5B, which beat DeepSeek baselines on shoestring coaching budgets.
If you happen to care about open ecosystems or on-premise choices, that is the 12 months China’s open-weight scene stopped being a curiosity and have become a critical various.
3. Small and native fashions develop up
One other factor I’m grateful for: we’re lastly getting good small fashions, not simply toys.
Liquid AI spent 2025 pushing its Liquid Basis Fashions (LFM2) and LFM2-VL vision-language variants, designed from day one for low-latency, device-aware deployments — edge bins, robots, and constrained servers, not simply large clusters. The newer LFM2-VL-3B targets embedded robotics and industrial autonomy, with demos deliberate at ROSCon.
On the big-tech aspect, Google’s Gemma 3 line made a powerful case that “tiny” can nonetheless be succesful. Gemma 3 spans from 270M parameters up by means of 27B, all with open weights and multimodal help within the bigger variants.
The standout is Gemma 3 270M, a compact mannequin purpose-built for fine-tuning and structured textual content duties — suppose customized formatters, routers, and watchdogs — lined each in Google’s developer weblog and neighborhood discussions in local-LLM circles.
These fashions could by no means pattern on X, however they’re precisely what you want for privacy-sensitive workloads, offline workflows, thin-client gadgets, and “agent swarms” the place you don’t need each software name hitting a large frontier LLM.
4. Meta + Midjourney: aesthetics as a service
One of many stranger twists this 12 months: Meta partnered with Midjourney as a substitute of merely making an attempt to beat it.
In August, Meta introduced a deal to license Midjourney’s “aesthetic expertise” — its picture and video technology stack — and combine it into Meta’s future fashions and merchandise, from Fb and Instagram feeds to Meta AI options.
VentureBeat lined the partnership in “Meta is partnering with Midjourney and can license its expertise for future fashions and merchandise,” elevating the plain query: does this sluggish or reshape Midjourney’s personal API roadmap? Nonetheless awaiting a solution there, however sadly, acknowledged plans for an API launch have but to materialize, suggesting that it has.
For creators and types, although, the instant implication is straightforward: Midjourney-grade visuals begin to present up in mainstream social instruments as a substitute of being locked away in a Discord bot. That would normalize higher-quality AI artwork for a a lot wider viewers — and pressure rivals like OpenAI, Google, and Black Forest Labs to maintain elevating the bar.
5. Google’s Gemini 3 and Nano Banana Professional
Google tried to reply GPT-5 with Gemini 3, billed as its most succesful mannequin but, with higher reasoning, coding, and multimodal understanding, plus a brand new Deep Assume mode for sluggish, onerous issues.
VentureBeat’s protection, “Google unveils Gemini 3 claiming the lead in math, science, multimodal and agentic AI,” framed it as a direct shot at frontier benchmarks and agentic workflows.
However the shock hit is Nano Banana Professional (Gemini 3 Professional Picture), Google’s new flagship picture generator. It makes a speciality of infographics, diagrams, multi-subject scenes, and multilingual textual content that really renders legibly throughout 2K and 4K resolutions.
On the planet of enterprise AI — the place charts, product schematics, and “clarify this method visually” photographs matter greater than fantasy dragons — that’s an enormous deal.
6. Wild playing cards I’m maintaining a tally of
Just a few extra releases I’m grateful for, even when they don’t match neatly into one bucket:
-
Black Forest Labs’ Flux.2 picture fashions, which launched simply earlier this week with ambitions to problem each Nano Banana Professional and Midjourney on high quality and management. VentureBeat dug into the main points in “Black Forest Labs launches Flux.2 AI picture fashions to problem Nano Banana Professional and Midjourney.”
-
Anthropic’s Claude Opus 4.5, a brand new flagship that goals for cheaper, extra succesful coding and long-horizon activity execution, lined in “Anthropic’s Claude Opus 4.5 is right here: Cheaper AI, infinite chats, and coding abilities that beat people.”
-
A gradual drumbeat of open math/reasoning fashions — from Mild-R1 to VibeThinker and others — that present you don’t want $100M coaching runs to maneuver the needle.
Final thought (for now)
If 2024 was the 12 months of “one large mannequin within the cloud,” 2025 is the 12 months the map exploded: a number of frontiers on the high, China taking the lead in open fashions, small and environment friendly methods maturing quick, and artistic ecosystems like Midjourney getting pulled into big-tech stacks.
I’m grateful not only for any single mannequin, however for the truth that we now have choices — closed and open, native and hosted, reasoning-first and media-first. For journalists, builders, and enterprises, that range is the actual story of 2025.
Joyful holidays and greatest to you and your family members!
