Is "miqu" a leak of the Mistral-medium LLM?

Mistral, an LLM company well known for open sourcing two of the currently most powerful open source models, also has an API with more powerful models, including mistral-medium.

Model weights were leaked yesterday on 4-chan under the name miqu, with the claim that they are a quantized version of mistral-medium. More details in all the links here.

Resolves yes if we get proof or extremely convincing evidence (I won't close without agreement by bettors on both sides) that miqu is mistral-medium, no if we get proof or extreme convincing evidence that it's not.

If we don't get either within one year, I'll resolve NA.

Get Ṁ600 play money
Sort by:

Mistral has confirmed that miqu is one of their models, but didn't directly say that it was mistral-medium, just a Llama 2 fine-tune they made.

This, in addition to the model's outputs' similarity to the API version of mistral-medium, and its multilingual abilities, make it seem very likely that it is mistral-medium.

I'd like to resolve this yes in 24 hours, if no one has objections, and no new information comes out by then.

@traders I'm feeling quite convinced that this is mistral-medium. I'd love to hear from no voters, in case anyone has objections to me resolving!

@Vergissfunktor I don't think that holds up based on the current information. Arthur Mensch specifically chose the words "an old model we trained" without mentioning mistral medium. It's better to wait for more clarity rather than resolving prematurely.

@Vergissfunktor I think, for example, that there is a strong chance that there was a llama 2 fine-tuning project at mistral, and the miqu model was an early output of that process when they were figuring out their fine-tuning recipe, and the mistral-medium model is a later iteration of that project. If this is the case then miqu and mistral-medium are different models.

predicts NO

@Vergissfunktor it is specifically not confirmed to be that model.

I voted on the basis that you did not name the market:

"Is "miqu" a leak of a Mistral LLM?"

predicts NO

@RemNi exactly

sold Ṁ103 of NO

@Vergissfunktor I noticed that I was carrying more of the NO volume than my confidence in the outcome.

So I have sold my position.

I still believe that confirmation has not occurred yet

@JosefMitchell @RemNi Thanks for responding. I'll not resolve yet. Both being different Llama 2 fine-tunes on the same data, more or less, sounds possible, and very hard to disprove.

bought Ṁ30 of NO

@JosefMitchell Buying back in with a sum more appropriate to my confidence

bought Ṁ20 of YES

resolves yes :))) - confirmation by Mistral themselves so let’s resolve

@Soli that doesn't mean it's mistral medium though right?

@Soli they're saying it's a model they fine-tuned from llama 2. So it could be mistral medium, or it could be a different model

@RemNi what is this "old model they trained and distributed openly", I wonder?

@RemNi I do agree with you, though – this could mean that mistral-medium is a Llama 2 fine-tune. We'd need confirmation that miqu is that fine-tune, still.

sold Ṁ313 of YES

@Vergissfunktor I don't think it can get better than this - anyways I sold my Yes ✌🏻and happy with the profits I made

good luck getting better evidence

@Soli congrats on the profits, I sold my no lol. Not sure how this market should resolve, there might not be any new info in the short term

@Soli That's fair – I did say we needed extremely convincing evidence, but I am convinced now. I'll comment to see what other people think, wait 24 hours for feedback, and if there are no contrary opinions or evidence, I'll resolve.

Here's some more pretty telling evidence:

bought Ṁ300 of YES

@Vergissfunktor The link doesn't work it seems.

@a2bb interesting, it seems to work for me on mobile. I'll try it on a laptop later

@a2bb hmmm, also worked when I typed it in to my laptop browser. Maybe try it on a different device

(oops, repeat, ignore)

So this is currently trading at 70%, which is odd given the comments say it already happened. If indeed Medium is their best model AND it went unleaked for 30 days I'll resolve this to true but I want to be sure on both...

@ZviMowshowitz Are you sure you're commenting on to right question? This is about whether miqu is actually mistral-medium, as the leaker claimed, not whether Mistral's best model was unleaked for 30 days.

@Vergissfunktor Oh weird, yeah I am not. Someone must have tagged me and I got confused.

More related questions