AI World War 1 Just Began: SHOCKING New AI (Tülu 3) Destroys DeepSeek & OpenAI!

In a shocking turn of events, a new player has just taken the lead in what we're calling AI World War I. Ai2's latest model, Tulu 3445B, has come out of nowhere and outperformed both controversial Deepseek's model and OpenAI's 01 on several major benchmarks. The AI wars began when Deepseek, a Chinese startup, released a model that could match or even surpass OpenAI's strongest offerings for free, sparking a fierce rivalry. Then Alibaba got in on the action with their Quen 2.5 model, upping the ante even further. To make things more intense, Microsoft and OpenAI accused Deepseek of stealing their tech, adding more drama to the war. But now, with 2 Lou 3 405B on the scene, it's clear the stakes have been raised and the competition is more heated than ever. This is getting really interesting and I can't wait to see where it goes next. So here's the rundown. AI2, short for the Allen Institute for AI, based in Seattle, is a non profit research organization. They're famous for doing cutting edge stuff in NLP and other AI research areas. This time they're stepping into the Spotlight by releasing Touluth 3.405B. The name 405B literally points to the insane scale of parameters the model uses. Bigger models often show better reasoning abilities, or so we've seen historically, and Toulu 3, 405 B definitely follows that trend. It was apparently trained using 256 GPUs in parallel, which shows how big this project was in terms of compute resources. Now, Toulu 3.405B is special not just because it's massive, but also because it's fully open source, unlike many powerful models out there. Everything you need to recreate this model, like training code, data and instructions, has been freely released and is permissively licensed. The folks at AI2 are calling to Lou 3.45B a big step for the US in showing that America can still be a leader in building top tier open source AI without always relying on big corporate labs. It's part of a bigger conversation about OpenAI research. You know how a lot of AI breakthroughs get locked behind closed APIs or proprietary code? Well, todo3.45B wants to break that pattern, but let's talk performance. How do we measure something like this? AI2 tested it on a bunch of popular benchmarks including pop QA, GSM K and math and more. These tests cover everything from knowledge, recall, and factual correctness to advanced reasoning, math word problems, coding tasks and instruction following and Internal evaluations to LO3.405B beat deep seq V3 and GPT4O on bunch of tasks. For instance on an evaluation set called Pop QA which has over 14,000 knowledge questions from Wikipedia to lo3.4 5B did better than not only those two models but also Meta's Llama 3.1 405B version. Pretty nuts right? Another place to lou 3.5B shines is Math Tasks. On a benchmark called GSM8K which is all about grade school level math problems, Thulu 345B actually snagged the highest performance among models in its class. That's always interesting because math is for many models quite challenging, especially if they haven't been explicitly fine tuned in that domain. If you want to try to lo345b out for yourself, AI2 has a chatbot web app that showcases it. They also posted code on GitHub and on hugging Face. So if you're a developer or a researcher, everything is there. You can test it, adapt it, or even combine it with your own data. It's all free and open now. Tulu345B didn't just show up overnight, it's part of the bigger Toulu 3 family which AI2 released in stages. Previously there were 8B and 70B parameter versions that were also instruction tuned to match or or even surpass some closed models. This newest foreign 5B iteration is basically a let's push the scale as far as we can with open weights and see what happens moment. One of the biggest reasons to LO III 405B stands out is how it was trained. AI2 used advanced post training approaches that combine supervised fine tuning, preference learning like dpo, which stands for Direct Preference preference Optimization, and a novel approach called reinforcement learning with verifiable rewards or rlvr. Essentially, RLVR gives the model tasks where the answers can be definitively checked for correctness, like math equations or certain constrained instructions. The model gets a reward if it produces the correct solution. This approach, especially at the 405B scale, seems to give to LO3 an extra boost in tasks that that demand accurate reasoning. They also emphasized verifiable outcomes for math problem solving and instruction following instead of just having a random reward function. Toulu3415b is rewarded only when the solution is verified by a process that can be automated. That's how it manages to get better at these tasks without drifting off into weird tangents. AI2 also mentioned a few technical challenges in scaling to Lou3415B because 405B parameters is a lot training, it required 32 nodes, 256 GPUs running in parallel. They also use specialized distributed frameworks to handle the job. Another big piece was synergy with VLLM for large batch inference, letting them do parallel training and inference efficiently. This is not your typical train it on a single cluster for a weekend time type of situation. It's seriously big Comparisons are always fun to loo 3.405b was pitted against deepseek v3, gpt4o, llama 3.1, 405b and naos hermes 3.405b on many tests, especially mmlu for knowledge pop qa and gsm8k to lou 3.405b held its own showing that open models can absolutely compete with the big proprietary ones. However, they do note that GPT4O remains slightly stronger on a few tasks with Toulu 3.405B sometimes tying or coming very close. The gap however, is not huge. Safety was also a highlight according to AI Twoulu 3.405B outperformed DeepSeek V3, Llama 3.1 and Noos Hermes 3 in multiple safety tests like refusing harmful or disallowed requests. That's huge when you consider that open source models often get flak for lacking robust content filters, but they claim to lo345b does a decent job thanks to specialized data curation and preference fine tuning. Speaking of data curation, Tulu345B's pipeline is extensive. They meticulously curated prompts and instructions from a variety of open data sets, some synthetic data, real user interactions from sets like Wildchat, and existing open instruction sets like Flan v2. They also had to do big time decontamination of their training data. They made sure the prompts for training wouldn't overlap with the test sets too much. That's always a big question. If you accidentally train on your test examples, you inflate your numbers artificially so they took steps to ensure fairness. Then we have the multi stage approach. So supervised fine tuning on carefully selected data for general skill building, direct preference optimization for aligning the model's answers to certain style or correctness preferences, and RLVR for tasks with verifiable correct outcomes. Combined that yields a pretty balanced model that can do math, reasoning, knowledge, recall, coding, general chat, and handled safety constraints fairly well. An interesting takeaway is how to lo3445b focuses on precise instruction following. The project includes tasks where the model must produce exactly three paragraphs or a certain word count or certain constraints. If it fails, it literally gets zero reward. That's a tough scenario, but it leads to a model more likely to follow instructions to the letter. So if your team wants a model that's good at following very strict constraints, to LO3.45b might be an exciting candidate. Overall, to LO3.45b's open source stance is a big statement. It stands in contrast to open models that only give partial code or partial weights. With To Lou 3, AI2 is releasing everything the training recipes, the preference data sets, the chat template, the final instructions and the code for each step. So if you want to replicate these results or push them even further, you can. For those of you wondering about usage, to Lou 3.405B can be found on Hugging Face with both the code and the instructions. If you are less technical, you might want to head to AI2's web demo and talk to the Toloo 3 405B chatbot. They basically said, hey, test it while it's hot, because you never know when a new flagship model might pop up to beat it. That's the nature of AI development. Right now. Things move super fast, and if you're curious about the name Tulu, it refers to a hybrid camel breed between a Bactrian camel and a dromedary. They use that as a metaphor for hybrid approaches, so that's kind of cool. All in all, to lo3.405B feels like a significant milestone for the open source community, especially stateside. AI2 is basically showing that it's possible to keep pushing boundaries for large language models in an openly accessible manner, letting researchers and developers tinker with everything from the model code to the training steps. By surpassing DeepSeek v3 in many benchmarks, Tolu 3, 405 B signals a nice competition in the global AI space, bridging the gap between academic openness and top tier performance. If you want to run your own experiments, there's never been a better time to get your hands on such a robust open model. That's it for today's breakdown. Tolu 3.405B is the talk of the AI town. It's massive, it's performing incredibly well on multiple tasks and it's thoroughly open. Definitely check it out and let me know what you think about how open source AI is evolving. If you found this helpful, make sure to subscribe and hit that like button. Thanks for watching, and I'll see you in the next one.