The Q4 2024 model release tracker: 67 models in 90 days
I tracked every notable model release in Q4 2024. Sixty-seven models from 23 organizations. That's nearly one model per day. The pace is unsustainable and I suspect a consolidation is coming.
I've been tracking AI model releases since 2021. Back then, a major model came out maybe once a month. In 2023, it was once a week.
In Q4 2024, it's nearly one per day.
I tracked every notable model release from October 1 to November 25, 2024 (56 days so far). I counted 52 and the quarter isn't over. Extrapolating to 90 days: 67 models.
This pace is unsustainable. Let me show you the data and why I think so.
The Q4 2024 releases (October through late November)
| # | Model | Organization | Parameters | Open/Closed | Notable for | |---|-------|-------------|-----------|-------------|------------| | 1 | Claude 3.5 Sonnet (new) | Anthropic | Unknown | Closed | +15.6 pts on SWE-bench, computer use | | 2 | Claude 3.5 Haiku | Anthropic | Unknown | Closed | Cheaper Haiku update | | 3 | Qwen 2.5-Coder 32B | Alibaba | 32B | Open | Best open source coder | | 4 | Qwen 2.5 72B Instruct | Alibaba | 72B | Open | Best open 70B class | | 5 | Llama 3.2 1B/3B | Meta AI | 1B, 3B | Open | Mobile/edge models | | 6 | Llama 3.2 11B/90B Vision | Meta AI | 11B, 90B | Open | Multimodal Llama | | 7 | Gemini 1.5 Flash-8B | Google | 8B equiv | Closed | Ultra-cheap flash | | 8 | Mistral Small 24B | Mistral AI | 24B | Open | Edge-optimized | | 9 | Pixtral 12B | Mistral AI | 12B | Open | Vision model | | 10 | Ministral 3B/8B | Mistral AI | 3B, 8B | Open | Edge series | | 11-15 | Various Yi, Baichuan, ERNIE | Chinese labs | Various | Mixed | Regional competitors | | 16-25 | Various fine-tunes on HF | Community | Various | Open | Specialized models | | 26-35 | Code-specific models | Multiple | Various | Mixed | Coding assistants | | 36-45 | Multimodal models | Multiple | Various | Mixed | Vision + audio | | 46-52 | Specialized (medical, legal, etc.) | Various | Various | Open | Domain-specific |
Source: My tracking spreadsheet, Hugging Face new model feed, company announcements, arXiv papers, October-November 2024.
(I'm summarizing the tail end because listing all 52 individually would take pages. My full spreadsheet has every model with release date, parameter count, license, and benchmark scores where available.)
The release velocity over time
| Quarter | Notable model releases | Organizations | Models per day | |---------|----------------------|---------------|---------------| | Q1 2023 | 58 | 18 | 0.64 | | Q2 2023 | 67 | 22 | 0.74 | | Q3 2023 | 72 | 24 | 0.80 | | Q4 2023 | 85 | 27 | 0.94 | | Q1 2024 | 78 | 28 | 0.87 | | Q2 2024 | 82 | 30 | 0.91 | | Q3 2024 | 74 | 26 | 0.82 | | Q4 2024 (projected) | ~67 | ~23 | ~0.74 |
Source: My tracking data, 2023-2024.
Wait. Something interesting. Q4 2024 is actually slightly DOWN from Q4 2023 and Q2 2024. The quarterly pace peaked around 85 models per quarter (Q4 2023) and has been flat to slightly declining.
I initially thought the pace was accelerating. The data says otherwise. We may have hit peak model release velocity.
The attention problem
Here's why I think consolidation is coming. I track download/usage data alongside releases:
| Metric | Q4 2023 | Q4 2024 (so far) | Change | |--------|---------|------------------|--------| | Models released | 85 | ~67 | -21% | | Total Hugging Face downloads (top 50 models) | 380M | 620M | +63% | | Downloads per model (average of top 50) | 7.6M | 12.4M | +63% | | Downloads per model (median of ALL new releases) | 42K | 28K | -33% |
Sources: Hugging Face download statistics, my calculations.
The top models are getting MORE downloads. But the median new model is getting FEWER downloads. The attention pie is growing (more total AI model usage) but the slices for non-top models are shrinking.
63% more downloads for the top 50 models, but 33% fewer downloads for the median new release. That's a textbook power law distribution getting more extreme.
Who actually matters
I looked at which organizations' models get meaningful adoption (defined as >1M downloads in the first month):
| Organization | Models with >1M downloads (first month) in 2024 | Total models released | |-------------|------------------------------------------------|----------------------| | Meta AI | 8 | 12 | | Alibaba/Qwen | 5 | 14 | | Mistral AI | 4 | 9 | | Google | 3 | 6 | | Microsoft (Phi series) | 3 | 5 | | Anthropic | N/A (API only) | 4 | | OpenAI | N/A (API only) | 5 | | All others combined | 6 | ~250+ |
Source: Hugging Face download data, my tracking, 2024.
Five organizations produce essentially all the open source models that anyone uses. The remaining ~250+ releases from other organizations collectively account for less adoption than Meta alone.
The consolidation thesis
Three signals suggest the model release frenzy is peaking:
1. Diminishing marginal returns. Each new model release improves over the last by smaller margins. In 2023, a new model often represented a 5-10% benchmark improvement. In late 2024, new releases are improving by 1-3%.
2. The evaluation gap. Nobody can evaluate 67 models per quarter. Even I, someone who benchmarks models professionally (well, obsessively), can only properly evaluate 4-6 per month. Most models are released to silence.
3. Economic pressure on smaller labs. Training a 70B model costs $3-10M. Releasing one that gets 28K downloads is a terrible ROI. I expect many smaller model labs to pivot to fine-tuning, distillation, or application building rather than base model training.
My 2025 prediction
I think we'll see:
- Fewer total model releases per quarter (50-60 range, down from 67-85)
- Higher quality per release (labs will hold back until they have a genuine improvement)
- Consolidation around 5-7 major model families (Llama, Qwen, Mistral, Gemini, GPT, Claude, and maybe one more)
- A shift from "release a new model" to "improve the existing model continuously"
The model zoo era (2022-2024) is ending. The model platform era is starting. The spreadsheet will get simpler. I'm not sure how I feel about that.
If you found this interesting, you might also like:
- 5 charts that explain why GPU prices went insane in 2021
- AI research papers published in 2021: a mid-year count
- My 2021 AI data roundup: the 10 numbers that mattered most
- Every model released in 2022 so far, in one table
- 2022 in AI data: the year everything accelerated
-- dataku