Resolves YES if a language model is released before January 1st 2026, that:
1. Has freely accessible weights, meaning the general public can download it and run it locally, regardless of additional restrictions.
2. Is explicitly described as having less than 10 billion parameters. (If the actual parameter number is less than 10B and was e.g. rounded up, this doesn't count as being under 10B.)
3. Achieves an Arena Score on http://lmarena.ai/leaderboard greater than the score of Claude 3.5 Haiku 2024-10-22, with both scores measured at the same point in time.
In the event that the way Arena Scores are calculated changes significantly or that specific Haiku model is no longer ranked before EOY2025 (I would be very surprised if this happens), I will try to find a suitable replacement criteria that traders can agree is fair. If no such criteria can be found, this market will N/A.
Update 2025-06-26 (PST) (AI summary of creator comment): The creator has announced their intention to resolve this market to YES in 24 hours, citing the release of Gemma 3n 4b. Please see the linked comment for their full reasoning.
1,000🏅 Top traders
| # | Name | Total profit |
|---|---|---|
| 1 | Ṁ875 | |
| 2 | Ṁ74 | |
| 3 | Ṁ62 | |
| 4 | Ṁ15 | |
| 5 | Ṁ10 |
People are also trading
@traders Gemma 3n e4b is finally fully released and available to download. Will resolve YES in 24 hours if there are no objections. (E.g, since the parameters are counted a little differently)
3 months later and we have an open model that's less than half of 10B and almost as good as Sonnet. I'm blown away. This will resolve YES as soon as it's possible to publicly access the weights, though I could be convinced to resolve it early if that's not controversial.
@MingCat lmsys is not assessing quality very well, at least. and llms at constant size are getting much better over time
@Bayesian Yeah, I'm hopeful too. Looks like Reka Core is 67B parameters. So the question is basically just how quickly we'll see small open-source models scale. We've seen AI development show some pretty weird progress overhangs where the tech theoretically should exist, but no one's properly capitalized on it. (It took a while for DeepSeek to come along, for instance)
@Bayesian looks like Gemma 3 is moving from 9b to 12b, otherwise it probably would have ended up qualifying
