💯 ThursdAI - 100th episode 🎉 - Meta LLama 4, Google tons of updates, ChatGPT memory, WandB MCP manifesto & more AI news

Hey Folks,

Alex here, celebrating an absolutely crazy (to me) milestone, of #100 episodes of ThursdAI 👏 100 episodes in a year and a half (as I started publishing much later than I started going live, and the first episode was embarrassing), 100 episodes that documented INCREDIBLE AI progress, we mention on the show today, we used to be excited by context windows jumping from 4K to 16K!

I want to extend a huge thank you to every one of you, who subscribes, listens to the show on podcasts, joins the live recording (we regularly get over 1K live viewers across platforms), shares with friends and highest thank you for the paid supporters! 🫶 Sharing the AI news progress with you, energizes me to keep going, despite the absolute avalanche of news every week.

And what a perfect way to celebrate the 100th episode, on a week that Meta dropped Llama 4, sending the open-source world into a frenzy (and a bit of chaos). Google unleashed a firehose of announcements at Google Next. The agent ecosystem got a massive boost with MCP and A2A developments. And we had fantastic guests join us – Michael Lou diving deep into the impressive DeepCoder-14B, and Liad Yosef & Ido Salomon sharing their wild ride creating the viral GitMCP tool.

I really loved today's show, and I encourage those of you who only read, to give this a watch/listen, and those of you who only listen, enjoy the recorded version (though longer and less edited!)

Now let's dive in, there's a LOT to talk about (TL;DR and show notes as always, at the end of the newsletter)

Open Source AI & LLMs: Llama 4 Takes Center Stage (Amidst Some Drama)

Meta drops Llama 4 - Scout 109B/17BA & Maverick 400B/17BA (Blog, HF, Try It)

This was by far the biggest news of this last week, and it dropped... on a Saturday? (I was on the mountain ⛷️! What are you doing Zuck)

Meta dropped the long awaited LLama-4 models, huge ones this time

* Llama 4 Scout: 17B active parameters out of ~109B total (16 experts).

* Llama 4 Maverick: 17B active parameters out of a whopping ~400B total (128 experts).

* Unreleased: Behemoth - 288B active with 2 Trillion total parameters chonker!

* Both base and instruct finetuned models were released

These new models are all Multimodal, Multilingual MoE (mixture of experts) architecture, and were trained with FP8, for significantly more tokens (around 30 Trillion Tokens!) with interleaved attention (iRoPE), and a refined SFT > RL > DPO post-training pipeline.

The biggest highlight is the stated context windows, 10M for Scout and 1M for Maverick, which is insane (and honestly, I haven't yet seen a provider that is even remotely able to support anything of this length, nor do I have the tokens to verify it)

The messy release - Big Oof from Big Zuck

Not only did Meta release on a Saturday, messing up people's weekends, Meta apparently announced a high LM arena score, but the model they provided to LMArena was... not the model they released!?

It caused LMArena to release the 2000 chats dataset, and truly, some examples are quite damning and show just how unreliable LMArena can be as vibe eval.

Additionally, during the next days, folks noticed discrepancies between the stated eval scores Meta released, and the ability to evaluate them independently, including our own Wolfram, who noticed that a quantized version of Scout, performed better on his laptop while HIGHLY quantized (read: reduced precision) than it was performing on the Together API inference endpoint!?

We've chatted on the show that this may be due to some VLLM issues, and speculated about other potential reasons for this.

Worth noting the official response from Ahmad Al-Dahle, head of LLama at Meta, who mentioned stability issues between providers and absolutely denied any training on any benchmarks

Too big for its own good (and us?)

One of the main criticism the OSS community had about these releases, is that for many of us, the reason for celebrating Open Source AI, is the ability to run models without network, privately on our own devices.

Llama 3 was released in 8-70B distilled versions and that was incredible for us local AI enthusiasts! These models, despite being "only" 17B active params, are huge and way to big to run on most local hardware, and so the question then is, if we're getting a model that HAS to run on a service, why not use Gemini 2.5 that's MUCH better and faster and cheaper than LLama?

Why didn't Meta release those sizes? Was it due to an inability to beat Qwen/DeepSeek enough? 🤔

My Take

Despite the absolutely chaotic rollout, this is still a monumental effort from Meta. They spent millions on compute and salaries to give this to the community. Yes, no papers yet, the LM Arena thing was weird, and the inference wasn't ready. But Meta is standing up for Western open-source in a big way. We have to celebrate the core contribution while demanding better rollout practices next time. As Wolfram rightly said, the real test will be the fine-tunes and distillations the community builds on these base models. Releasing the base weights is crucial for that. Let's see if the community can tame this beast once the inference dust settles. Shout out to Ahmed Al-Dahle and the whole Llama team at Meta – incredible work, messy launch, but thank you for pushing open source forward. 🎉

Together AI & Agentica (Berkley) finetuned DeepCoder-14B with reasoning (X, Blog)

Amidst the Llama noise, we got another stellar open-source release! We were thrilled to have Michael Lou from Agentica/UC Berkeley join us to talk about DeepCoder-14B-Preview which beats DeepSeek R1 and even o3-mini on several coding benchmarks.

Using distributed Reinforcement Learning (RL), it achieves 60.6% Pass@1 accuracy on LiveCodeBench, matching the performance of models like o3-mini-2025-01-31 (Low) despite its smaller size.

The stated purpose of the project is to democratize RL and they have open sourced the model (HF), the dataset (HF), the Weights & Biases logs and even the eval logs!

Shout out to Michael, Sijun and Alpay and the rest of the team who worked on this awesome model!

NVIDIA Nemotron ULTRA is finally here, 253B pruned Llama 3-405B (HF)

While Llama 4 was wrapped in mystery, NVIDIA dropped their pruned and distilled finetune of the previous Llama chonker 405B model, turning at just about half the parameters.

And they were able to include the LLama-4 benchmarks in their release, showing that the older Llama, finetuned can absolutely beat the new ones at AIME, GPQA and more.

As a reminder, we covered the previous 2 NEMOTRONS and they are a combined reasoning and non reasoning models, so the jump is not that surprising, and it does seem like a bit of eval cherry picking happened here.

Nemotron Ultra supports 128K context and fits on a single 8xH100 node for inference. Built on open Llama models and trained on vetted + synthetic data, it's commercially viable. Shout out to NVIDIA for releasing this, and especially for pushing open reasoning datasets which Nisten rightly praised as having long-term value beyond the models themselves.

More Open Source Goodness: Jina, DeepCogito, Kimi

The open-source train didn't stop there:

* Jina Reranker M0: Our friends at Jina released a state-of-the-art multimodal reranker model. If you're doing RAG with images and text, this looks super useful for improving retrieval quality across languages and modalities (Blog, HF)

* DeepCogito: A new company emerged releasing a suite of Llama fine-tunes (3B up to 70B planned, with larger ones coming) trained using a technique they call Iterated Distillation and Amplification (IDA). They claim their 70B model beats DeepSeek V2 70B on some benchmarks . Definitely one to watch. (Blog, HF)

* Kimi-VL & Kimi-VL-Thinking: MoonShot, who sometimes get lost in the noise, released incredibly impressive Kimi Vision Language Models (VLMs). These are MoE models with only ~3 Billion active parameters, yet they're showing results on par with or even beating models 10x larger (like Gemma 2 27B) on benchmarks like MathVision and ScreenSpot. …