AI Roundup April 2024: Llama 3 Drops, Mixtral 8x22B and Phi-3 Crowd the Open Weights

TL;DR

April 2024 was the month open weights got serious. Meta shipped Llama 3 in 8B and 70B sizes, Mistral dropped the Mixtral 8x22B MoE under Apache 2.0, and Microsoft put out Phi-3, a 3.8B model that punches like something three times its size. On the product side, Adobe wired generative video (including OpenAI's Sora) into Premiere Pro, and Microsoft made a $1.5B geopolitical bet on Abu Dhabi's G42. If you build with local models, this was your month.

Meta Ships Llama 3 8B and 70B, and the Local Stack Reorganizes Around It

On April 18, Meta released Llama 3 with pretrained and instruction-tuned variants at 8B and 70B parameters. The headline is that benchmarks jumped hard without the model getting bigger. The lever was data: Meta trained on roughly 15 trillion tokens, far more than Llama 2 saw, and it shows in reasoning, coding, and instruction following. Meta's own numbers had Llama 3 70B beating Gemini Pro 1.5 and Claude 3 Sonnet on a chunk of standard benchmarks.

Why it matters for builders: the 8B is the new default for local. It runs comfortably on a single consumer GPU, quantizes well, and is good enough that a lot of people stopped reaching for a closed API for everyday tasks. The 70B is the new self-hosted workhorse for anyone with the VRAM. Within days the model was in Ollama, llama.cpp, and every inference stack worth using, and GGUF quants were everywhere.

Meta also flagged a 400B+ model still in training, which set the expectation that an open model would soon be sniffing at GPT-4 class territory. The license is the usual caveat: permissive enough for almost everyone, but not OSI open source, with the over-700M-monthly-active-users clause aimed squarely at the other hyperscalers.

Mistral Drops Mixtral 8x22B Under Apache 2.0

The day before Llama 3, on April 17, Mistral released Mixtral 8x22B, a sparse mixture-of-experts model with 141B total parameters but only about 39B active per token. That is the whole pitch of MoE: you pay compute like a ~39B model while carrying the knowledge of something much larger. It shipped under a genuine Apache 2.0 license, which is the part that matters if you are building a commercial product and do not want to read a custom license with a lawyer.

It is heavy to run, you need real hardware or aggressive quantization to fit it, but the efficiency-per-active-parameter story and the clean license made it an instant favorite for teams that want frontier-ish capability they fully control. Two of the strongest open releases of the year landed within 24 hours of each other, which tells you how fast this race had gotten.

Microsoft's Phi-3 Proves Small Models Got Scary Good

On April 23, Microsoft introduced the Phi-3 family, led by Phi-3-mini at just 3.8B parameters, with Small (7B) and Medium (14B) following. The claim, backed up reasonably well in practice, is that Phi-3-mini trades blows with Llama 3 8B and Mistral 7B while using a fraction of the compute. The trick was not scale, it was data quality: a heavily curated, textbook-style training set.

For anyone building on edge or constrained hardware, this is the interesting thread of the month. A 3.8B model that is actually useful runs on a phone, a laptop, or a Raspberry-class device. Phi-3-mini showed up on Azure AI Studio, Hugging Face, and Ollama right away. The lesson April hammered home: parameter count is not the whole story, and small-but-clean is a real strategy.

Adobe Brings Generative Video to Premiere Pro, Including Sora

On April 15, Adobe previewed generative AI video tools inside Premiere Pro, powered by a new Firefly video model. The built-in features include Generative Extend (add frames to lengthen a clip), plus object addition and removal with smart masking, and text-to-video.

The headline move was openness: Adobe said it was in early explorations to let editors plug in third-party models directly in the timeline, naming OpenAI's Sora, Runway, and Pika. That matters because it reframes generative video from a standalone toy into a normal step in a professional workflow, with Content Credentials attached so you can see which model touched a clip. Sora itself was still not public, but seeing it slotted into the industry-standard NLE put real weight behind the it-is-coming narrative.

Microsoft Puts $1.5B Into Abu Dhabi's G42

On April 16, Microsoft announced a $1.5 billion strategic investment in G42, the UAE AI holding company, with Microsoft's Brad Smith taking a board seat. This was less about a model and more about where AI infrastructure and influence flow. As part of the deal, G42 agreed to pivot away from Chinese technology toward American stacks, and the arrangement came wrapped in a first-of-its-kind intergovernmental assurance agreement on AI safety and security.

For builders this is background radiation, but it is the kind that shapes the ground you build on: who controls compute, which clouds get the chips, and how AI capability gets exported across borders. April was the month that conversation got concrete dollars attached.

Key Takeaways

• Open weights came of age: Llama 3, Mixtral 8x22B, and Phi-3 all landed within a week, giving self-hosters genuinely strong options across every size class.
• Llama 3 8B is the new local default: it runs on one consumer GPU, quantizes cleanly, and is good enough to replace a closed API for a lot of everyday work.
• Licenses are a feature: Mixtral 8x22B's Apache 2.0 makes it the safer pick for commercial products versus Llama 3's custom, hyperscaler-targeted license.
• Small models got scary good: Phi-3-mini at 3.8B shows curated data beats raw scale for edge and on-device use.
• Generative video entered real workflows: Adobe putting Firefly and Sora inside Premiere Pro signals the move from demo to production tool.
• The geopolitics got expensive: Microsoft's $1.5B G42 deal shows AI compute and influence are now a cross-border chess game, not just a model race.

Sources: Meta: Introducing Llama 3, Mistral: Mixtral 8x22B, Microsoft: Introducing Phi-3, Adobe: Gen AI in Premiere Pro, Microsoft: $1.5B investment in G42