April 2024 wasn’t about a single frontier launch. It was about a structural shift: in the span of two weeks, open-weight and small models reached a quality bar that, until then, only sat behind a paid API. The practical question for most teams stopped being “which closed model do we call?” and started being “what can we now run, fine-tune, or self-host that we couldn’t last quarter?” That changes the build-vs-buy math, and it changes it in your favor.
Meta Ships Llama 3 — and Open Weights Catch Up
On April 18, Meta released Llama 3 in 8B and 70B sizes under a permissive commercial license, trained on roughly seven times the data of Llama 2. The 70B traded blows with frontier closed models on key benchmarks; the 8B punched far above its size.
Why it matters for operators: This is the month “open” stopped meaning “compromise.” If your use case involves sensitive data, predictable per-token economics, or latency you control, you now have a credible model you can host yourself. The 8B in particular is the workhorse most teams overlook — good enough for classification, extraction, and routing at a fraction of frontier cost. Don’t reach for the biggest model by reflex.
Mixtral 8x22B and Phi-3: Two Bets on Efficiency
Days earlier, on April 17, Mistral released Mixtral 8x22B under Apache 2.0 — a sparse mixture-of-experts model with 141B total parameters but only ~39B active per token. Then on April 23, Microsoft released Phi-3-mini, a 3.8B model trained on curated data that benchmarked near GPT-3.5 and is small enough to run on a phone.
Why it matters for builders: These are two different answers to the same question — how do you get strong output without paying for strong compute? Mixtral keeps capability high while cutting inference cost. Phi-3 proves that data quality, not raw size, drives a lot of real-world performance. For a growth-stage team, the lesson is concrete: a well-chosen small model plus good retrieval and prompting often beats a giant model used lazily. Cheaper, faster, and easier to own.
OpenAI Makes GPT-4 Turbo with Vision Generally Available
On April 9, OpenAI made GPT-4 Turbo with Vision generally available via its API (the gpt-4-turbo-2024-04-09 snapshot), with a 128K context window and multimodal input in the mainline model.
Why it matters: Vision quietly became a default capability, not a separate endpoint. That unlocks document parsing, screenshot-driven workflows, and UI understanding without bolting on a second system. If you’ve been treating image input as a someday feature, the plumbing is now boring enough to build on today.
The Money Followed: a16z’s $7.2B Raise
In April, Andreessen Horowitz closed $7.2 billion in new funds, with explicit allocations toward AI infrastructure and AI applications. Capital at that scale signals where the smart money expects durable value to accrue.
Why it matters: Funding headlines don’t change your roadmap, but they do tell you the ecosystem around you — tooling, talent, infrastructure — will keep maturing fast. Build on abstractions you can swap. The model layer is commoditizing in real time, and April was the proof.
The Through-Line
April 2024’s signal was clear: model choice is now a portfolio decision, not a default. Open weights are viable, small models are capable, and the cost curve is bending hard. The teams that win won’t be the ones using the most expensive model — they’ll be the ones matching the right model to each job and owning the parts that matter.
If you’re weighing what to self-host, what to buy, and what to skip, that’s exactly the kind of call we help founders make. Let’s talk.