AI model release frequency by quarter: a 4-year chart

I've been counting notable AI model releases every quarter since Q1 2021. This quarter, something changed.

The count went down.

Quarterly model releases (Q1 2021 to Q2 2025)

| Quarter | Notable releases | Quarter-over-quarter change | |---------|-----------------|---------------------------| | Q1 2021 | 8 | - | | Q2 2021 | 10 | +25% | | Q3 2021 | 11 | +10% | | Q4 2021 | 9 | -18% | | Q1 2022 | 14 | +56% | | Q2 2022 | 18 | +29% | | Q3 2022 | 21 | +17% | | Q4 2022 | 24 | +14% | | Q1 2023 | 31 | +29% | | Q2 2023 | 38 | +23% | | Q3 2023 | 42 | +11% | | Q4 2023 | 51 | +21% | | Q1 2024 | 58 | +14% | | Q2 2024 | 63 | +9% | | Q3 2024 | 67 | +6% | | Q4 2024 | 89 | +33% | | Q1 2025 | 78 | -12% | | Q2 2025 | 54 | -31% |

Sources: My model tracking spreadsheet, Hugging Face, arXiv, Papers With Code, Epoch AI. "Notable" = models with technical reports or from established labs.

Q4 2024 was the peak: 89 notable releases. That was the DeepSeek V3 wave, the Llama 3.3 variants, and a surge of reasoning model experiments.

Q1 2025 dropped to 78 (-12%). Q2 2025 dropped further to 54 (-31%). The first consecutive quarterly decline in the dataset.

What happened?

| Factor | Impact on release count | |--------|----------------------| | MoE transition | Labs spending more time on fewer, larger MoE models | | DeepSeek R1 reset expectations | Many labs paused to incorporate reasoning training | | Funding tightening | Smaller labs running out of compute budget | | Diminishing returns on fine-tuning | Community fine-tunes adding less value over better base models | | Benchmark saturation | Less motivation to release when MMLU is saturated |

The biggest factor is probably the MoE transition. Building a mixture-of-experts model takes longer than a dense model. Labs that were releasing a new model every 2-3 months are now spending 4-6 months per release.

Release count by organization type

| Organization type | Q4 2024 | Q2 2025 | Change | |-------------------|---------|---------|--------| | Major labs (FAANG+) | 12 | 10 | -17% | | Funded startups | 18 | 14 | -22% | | Chinese labs | 15 | 12 | -20% | | Community/individual | 44 | 18 | -59% |

Sources: My categorization of model releases.

Community and individual model releases dropped 59%. That's the sharpest decline. The base models have gotten so good that fine-tuning for marginal improvements is less worthwhile.

Major labs dropped only 17%. They're still releasing steadily, just with longer intervals between releases.

The growth rate story

| Year | Total releases | Year-over-year growth | |------|---------------|---------------------| | 2021 | 38 | - | | 2022 | 77 | +103% | | 2023 | 162 | +110% | | 2024 | 277 | +71% | | 2025 (H1, annualized) | 264 | -5% (est.) |

Growth went from +110% to +71% to -5%. The S-curve is bending.

My interpretation

This isn't a crisis. It's maturation. The model release frenzy of 2023-2024 was partly driven by "gold rush" energy: every lab wanted to be on the leaderboard, every researcher wanted to publish a new model.

Now the value is shifting from "release a model" to "deploy a model in production." Building reliable agent systems, optimizing inference costs, and creating useful products is where the energy is going.

I'm not sad about this. Keeping up with 89 models per quarter was exhausting. 54 is manageable. I can actually evaluate them properly instead of just noting their existence.

My quarterly model count chart has a new shape. Four years of going up and to the right. And now, for the first time, a bend.

Peak model. I think we're past it.

---

If you found this interesting, you might also like:

• 5 charts that explain why GPU prices went insane in 2021
• I counted every AI model released this quarter. Here's what I found.
• AI research papers published in 2021: a mid-year count
• My 2021 AI data roundup: the 10 numbers that mattered most
• I tracked AI image generation quality over 6 months. The improvement rate is scary.

-- dataku