[December 2025] AI & Machine Learning Monthly Newsletter 💻🤖

72nd subject! In the event you missed them, you’ll be able to read the previous issues of my monthly A.I. & Machine Learning newsletter here.

Hey everybody!

Daniel right here, I’m a machine studying engineer who teaches the next beginner-friendly machine studying programs:

I additionally write repeatedly about machine studying alone weblog in addition to make movies on the subject on YouTube.

Since there’s lots happening, the utmost care has been taken to maintain issues to the purpose.

This is what you might need missed in December 2025 as an A.I. & Machine Studying Engineer… let’s get you caught up!

My work

Joyful New Yr everybody!

I hope you’re as excited for 2026 as I’m.

Regardless of being stuffed with holidays, the final month of 2025 didn’t change something for machine studying and AI updates… there’s fairly a couple of!

Just a few notes alone work after which we’ll get into them.

• My new course is dwell! The Machine Learning with Hugging Face Bootcamp is live on Zero to Mastery. You possibly can consider Hugging Face because the homepage for contemporary ML and AI workflows. I personally use Hugging Face day by day in each an exploratory {and professional} sense. This new course goals to rebuild the precise workflows I take advantage of for creating customized fashions and dealing with shoppers in addition to particular ML and AI tasks. Inside we’ll go step-by-step by means of a number of ML tasks, ending with sharable demos on the finish.
• [Video] Ranking the best open-source AI companies and models of the year. Writing these ML Month-to-month points means I get to see and work together with fairly a couple of fashions and open-source AI releases. On this video, I am going by means of my high corporations and releases of the 2025.
• [Video] Unboxing + setting up the NVIDIA DGX Spark. NVIDIA determined to ship me one among their new DGX Sparks (a small AI supercomputer). Keep tuned for extra movies exploring how this machine works for native AI growth and use.

From the Web

Instance outcomes of utilizing Tinker for fine-tuning a Qwen3-VL mannequin for picture classification versus a pure vision-based mannequin DINOv2. Discover how the Qwen3-VL mannequin drastically improves with a small variety of samples. Supply: Pondering Machines weblog.

• Apple show how the M5 chip’s neural accelerators speedup LLM generation. In comparison with the M4 chip within the MacBook Professional, the M5 chip can obtain as much as 4x sooner Time to First Token (TTFT) on fashions resembling Qwen3-14B-MLX-4bit. It additionally achieves a mean of 25% sooner era velocity due to the elevated reminiscence bandwidth. I feel we’re solely beginning to scratch the floor of what’s doable operating native fashions. This highlights how a lot can enhance in a single era of {hardware}.

Examples of how a lot the M5 chip hurries up Time to First Token (TTFT) compared to the M4 chip throughout varied mannequin architectures, usually averaging 3-4x enchancment. Supply: Apple Machine Studying weblog.

• AI World Shares their 2025 AI Advent Calendar. My favorite was the highest downloaded fashions of the yr. One thing non-surprising was seeing the Qwen fashions come out on high for probably the most variety of downloads. In response to the web site, the Qwen2.5-VL-3B-Instruct mannequin has been downloaded greater than 300,000,000 instances 😲. Although I’m unsure what counts as a obtain… is it at all times a contemporary set up? In that case, that’s a a lot bigger quantity than I believed.

• [Essay] Why your boss isn’t worried about AI by Boyd Kane. Thought frightening piece on how AI fashions generally get misinterpreted as common software program and so the potential downsides of AI get ignored as a result of we largely perceive the downsides of errors in software program. I just like the recurring theme all through of: it largely comes all the way down to what was the info the AI was educated on?
• Google Colab comes to VS Code. No native GPU? No downside. Now you can join a Google Colab backend with a cloud-hosted GPU and run your native notebooks on Google Colab straight from VS Code. This implies you possibly can write experimentation code domestically after which when it comes time to run it on a GPU, you possibly can connect with Google Colab and make it run sooner.
• Google published a nice review of all their releases in 2025 (there’s fairly a couple of). Sitting behind OpenAI by way of uncooked mannequin efficiency for 3 years (since GPT-4 launched), now arguably on high by virtually each metric. Google has been on an absolute roll. I’m personally utilizing Gemini way more on the finish of 2025 than in the beginning.
• Google DeepMind release WeatherNext 2. Based mostly on a brand new Purposeful Generative Community (FGN), the WeatherNext 2 mannequin outperforms the earlier era on 99.9% of measurable metrics together with temperature and wind predictions in addition to lead instances (0-15 days).
• Philipp Schmid releases a collection of weblog posts to assist with Gemini 3 and AI Agent Understanding. From Gemini 3 Prompt Best Practices, Why Senior Engineers Struggle to Build AI Agents (a very good overview of how generally extra data can maintain you again when studying a brand new idea), A Practical Guide on Building an Agent from Scratch with Gemini 3 and Context Engineering for AI Agents (on of my favorite takeaways was the purpose on as AI fashions get higher and higher, the necessity for extreme directions and harnesses will get eliminated).
• The Hugging Face Tokenizers library gets several upgrades as part of Transformers v5. Tokenization is the method of turning uncooked information into numbers (e.g. textual content into numerical type so a machine studying mannequin can work with it). And as a part of the transformers v5 launch, the tokenizers library (which is built-in to transformers) will get upgrades together with a speedy Rust-based tokenizer by default and less complicated methods to create and customise your personal tokenizers. Bonus: Tokenizer behaviour may be difficult as a result of not all tokenizers are made the identical. And in case you use the unsuitable tokenizer for the unsuitable mannequin, you will get lower than very best outcomes. See the *Gotchas in Tokenizer Behavior* article for extra.
• Beej’s Guides (by Brian Hall). Once in a while you stumble an distinctive weblog somebody has been toiling away for on for years. And this month’s for me is beej.us. There are some nice guides on there I’d counsel trying out in case you wish to learn HTML-style tech blogs and simple posts (one among my favorite types). I’ve been studying the one on Git and plan on studying the one on Computer Science subsequent. There’s additionally an upcoming one on Python programming.
• [Essay] The Future of Software Development is Software Developers by Jason Gorman. Jason Gorman has been a pc programmer for 43 years, seeing traits coming and going. He writes that LLMs aren’t any completely different. And argues that not solely will LLMs not essentially exchange human programmers, they’ll demand extra. Jevon’s paradox in full swing. I’m seeing the identical on the bottom in my very own work, now that extra prototypes can be created, the urge to create them is even larger. However as at all times, a prototype isn’t a completed product. So ultimately, your job is to still deliver code you have proven to work.
• [Guide] How to speedup open-source LLMs with EAGLE-3 by Lmsys.org. EAGLE-3 (Extrapolative Attention Guided LEarning) is a speculative decoding method that hurries up LLM decoding velocity by 2-3x by including a small ‘draft head’ layer to the mannequin which is round 2-5% of the mannequin’s whole parameters. The draft head generates candidate tokens (quick) and the massive mannequin selects the most effective (or it generates new tokens if the draft tokens don’t match the anticipated outputs).

Instance of EAGLE 3 in use. The unique mannequin is fitted with an EAGLE 3 head which is used for producing a tree of draft output tokens which may be chosen by the bigger mannequin for remaining outputs. Supply: Lmsys weblog.

Daniel’s Open-Supply AI of the Month

• Meta release OmniLingual, an open-source Automated Speech Recognition mannequin able to performing speech recognition for 1,600+ languages which outperforms Whisper v3. There are a number of mannequin sizes obtainable starting from 300M parameters to 9B.
• Ai2 release Molmo2, a collection of open-source VLMs (Imaginative and prescient Language Fashions) with a concentrate on openness, multi-image and video capabilities. For instance, the mannequin is extremely able to pointing to things in photos based mostly on a textual content immediate resembling “level to all of the seafood objects”. It may additionally monitor objects and objects in movies over a collection of frames. The discharge not solely comes with open mannequin weights, all the information is offered as properly.

Instance of Molmo2’s pointing capabilities. Given a picture an directions on what to level at, it’s able to returning text-based level coordinates which may be transformed to image-based level coordinates and proven on a plot. Supply: Allen AI playground with Molmo2 8B mannequin.

• EssentialAI release rnj-1 , a 8B LLM on par with fashions resembling Qwen3-8B. Attention-grabbing that the mannequin was educated with JAX on TPUs and AMD GPUs.
• Meta release SAM Audio, a segmentation mannequin which has been designed to separate audio from visible feeds. For instance, if there’s a video of an individual strolling alongside a road speaking on the telephone and a canine is barking within the background, you’ll be able to immediate the mannequin to pick the “individual” and the separate the audio tracks to spotlight the individual speaking an ignore the background noise.
• ServiceNow-AI releases Apriel-1.6-15b-Thinker, an open-source VLM which is on par or higher than Gemini 2.5 Flash in addition to Claude Haiku 4.5 on the Artificial Analysis Index (it scores 57 versus 54 for Gemini 2.5 Flash). Learn the discharge blog post for extra particulars.
• Z.ai releases GLM-4.6V-Flash, GLM-4.6V and GLM-4.7 (present greatest open-source coding mannequin). The 2 V fashions allow imaginative and prescient inputs and are very aggressive with different fashions at their measurement. The GLM-4.7 mannequin brings an unimaginable leap ahead for open-source fashions in coding capabilities. It performs on par with Claude Sonnet 4.5 in addition to GPT 5.1 (excessive) on a number of software-related benchmarks.
• Meta releases Segment Anything 3 (SAM 3), a mannequin able to segmenting and detecting objects in photos based mostly on textual content and picture immediate inputs. Bonus: See EfficientSAM3 for an environment friendly implementation with a 99% imaginative and prescient encoder and 6x smaller textual content encoder.
• CLaRA is an LLM from Apple which mixes retrieval and era right into a single mannequin. For instance, it dually trains the mannequin to study to compress queries and paperwork so when it comes time for RAG (Retrieval Augmented Era) the retrieval step has already been in-built.
• The Qwen staff releases a collection of picture mannequin updates. Qwen-Image-Layered (flip a picture into layers) takes an current layered picture and breaks it into particular person layers to allow them to be altered or edited individually. Qwen-Image-Edit-2511 (edit a picture based mostly on a textual content immediate) improves the consistency of picture enhancing over the earlier 2509 model. And Qwen-Image-2512 (generate a picture based mostly on a textual content immediate) improves the human realism, pure particulars and textual content rendering of the earlier model.
• GLiNERv2 is a text classification, entity extraction and structured data model all in one. For instance, you’ll be able to put in a string of textual content resembling “Daniel Bourke is a Machine Studying Engineer who builds Nutrify, an app for serving to folks find out about complete meals” and inform it to extract individual , job , firm and it’ll output Daniel Bourke, Machine Studying Engineer, Nutrify. The fashions are quick and are capable of run on CPU with GPU utilization providing much more speedups. See the instance beneath for utilization.

from gliner2 import GLiNER2


extractor = GLiNER2.from_pretrained("fastino/gliner2-base-v1")


textual content = "Apple CEO Tim Cook dinner introduced iPhone 15 in Cupertino yesterday."
consequence = extractor.extract_entities(textual content, ["company", "person", "product", "location"])

print(consequence)

• Encoder-only Mask Transfer (EoMT) is a plain Vision Transformer (ViT) able to producing segmentation masks at encoder-only speeds. Usually with a segmentation mannequin you might need a number of completely different elements to supply output segmentation masks. Nevertheless, since EoMT solely makes use of the ViT spine to supply output courses and masks, it might probably obtain related outcomes to extra advanced fashions with a a lot sooner inference time (as much as 4x sooner). Learn the paper, get the models on Hugging Face.

Instance of utilizing EoMT weights from Hugging Face to generate masks on a picture of meals objects. Meals picture generated with Gemini based mostly on obtainable meals courses within the COCO dataset. The underside determine is from the EoMT paper.

• PleAIs releases Baguettotron, Monad Small LLMs and SYNTH a synthetic dataset trained with seed documents. Baguettotron is a 321M parameter mannequin and Monad is a 50M parameter language mannequin. Every performs extremely properly for its measurement in comparison with different fashions resembling Gemma-3-270M and Qwen-3-600M. They have been each educated on the SYNTH dataset which is a reasoning centered dataset of 200B tokens created from 50,000 Wikipedia paperwork as seed inputs. It is a actually thrilling course for SLMs (small language fashions), these are the sorts of fashions that could possibly be in every single place doing particular duties with minimal compute footprint. Learn the blog announcement for extra.
• Ai2 release Olmo3 and Olmo3.1 LLM models. These fashions are on par with Qwen3-32B-Instruct however are utterly open from information to coaching code and methodologies.
• Mistral release the Mistral 3 series of models. These multi-modal (picture and textual content) fashions vary from 3B, 8B, 14B (Ministral 3) and 675B parameters (Mistral Giant 3). Mistral Giant 3 performs on par or higher than different massive open-source fashions resembling Kimi-K2 and DeepSeek 3.1. And the smaller fashions carry out on par or higher than the similar-sized Qwen3-VL variants. Notably the Ministral3 8B Instruct mannequin outperforms the bigger Gemma3 12B Instruct mannequin by a major margin. All fashions can be found underneath the Apache 2.0 license.
• Google release FunctionGemma-270M, a mannequin designed to be fine-tuned to name particular features. For instance, you possibly can think about an in-car assistant which is designed to alter settings in a automotive, you possibly can inform it to “modify the air conditioner to 22C” and the FunctionGemma-270M mannequin may name the “adjust_air_conditioner” perform.
• NVIDIA release Nemotron 3 Nano, a 31.6B whole parameter with ~3.6B lively parameter LLM mannequin which has best-in-class reasoning accuracy and is 4x sooner than Nemotron Nano 2 in addition to as much as 3.3x sooner than fashions resembling Qwen3-30B-A3B-Pondering-2507. The mannequin is prepared for deployment underneath industrial settings due to NVIDIA’s open mannequin license. Learn the blog post for more information.
• Amazon release Chronos-2, a common time collection forecasting mannequin. The mannequin is able to utilizing in-context studying for fixing forecasting duties with an infinite variety of dimensions in a zero-shot method. Learn the blog post announcement for extra info.
• ByteDance releases Dolphin-v2 for advanced document parsing. The mannequin is able to extracting 21 factor classes from a doc together with part headings, paragraphs, figures, captions, lists, watermarks, figures and extra. The doc extraction occurs in two phases, classification and format evaluation adopted by particular content material parsing (this two stage course of allows focused extraction for photographed and digital type paperwork).

Analysis

• VL-JEPA explores using Joint Embedding Predictive Structure (JEPA) for the vision-language house. Outcomes present that VL-JEPA will get related or higher outcomes on video classification and retrieval datasets when in comparison with SigLIP2 and Notion Encoder.
• The Qwen3-VL technical report was launched and its a treasure path of tidbits on the way to prepare a world-class open-source VLM.
• AnyUp is a technique to scale up the options of any imaginative and prescient characteristic at any decision. It may run at inference time and doesn’t require fine-tuning for a selected encoder. See the example notebook to run the demo by yourself photos.

Instance of utilizing AnyUp to upscale the output options of a DINOv2 mannequin.

Releases

Movies

See you subsequent month!

What an enormous month for the ML world in December and 2025 as properly! I am excited for 2026 and need you all a Joyful New Yr.

As at all times, let me know if there’s something you assume ought to be included in a future put up.

Within the meantime, continue to learn, maintain creating, maintain dancing.

See you subsequent month,

Daniel

www.mrdbourke.com | YouTube

By the way in which, I am additionally an teacher with Zero To Mastery Academy instructing folks Machine Studying & AI in probably the most environment friendly method doable. You possibly can see a couple of of our programs beneath or try all Zero To Mastery courses.