Mistral AI launches Mixtral-Next (opens in new tab)

Company creates blackbox technology, and the company's communications are themselves like a blackbox... fitting

(I know that Mistral does a lot more stuff in the open than other companies, just couldn't resist the parallel between this and the blackbox limitations of LLMs in general)

> for releasing new models is extremely low-information

To be fair, this is not a release. This was the previous release https://mistral.ai/news/mixtral-of-experts/

It looks more like not trying very hard to hide things until release, rather than being a black box.

From limited experimentation, and within the confines of a single prompt (rather than a full chat), this model seems reasonably interesting. Does anyone have good examples of QA that showcase the capabilities compared to other advanced models?

I only get the message "Connection errored out."

And how does Mistral do "accurate long content retrieval"?

I know what they were going for with the Mixtral name but every time I come across it I wonder if they considered just how easily the two might be confused. It seems like a poor branding decision - what if some expected the Mixtral performance but accidentally uses a Mistral model? What if someone wants the low resource usage of e.g. Mistral 7B but tries out Mixtral 8x7B instead? It's especially hard when your colleagues aren't necessarily native English speakers.

There's got to be a better name for such a cool product. Maybe MistralX? MistMix?

I’m pretty new to running these locally, but here’s my understanding:

Best models currently: codellama or deepseek coder. 6.7B or 1B depending on how much latency you can tolerate

Treesittee: from looking at the logs of the chat completions requests for Continue or Twinny extensions for VS Code, they both appear to just send a chunk of the document as well as a special placeholder to indicate where the cursor currently is.

Another one is https://github.com/WisdomShell/codeshell/blob/main/README_EN... and it has its own IntelliJ plugin.

I'm also interested in the answer to that.

> do coding LLMs use treesitter...?

It's up to the app to put that into the context. Generally, coding LLMs do well if you provide them the source tree, graph, search results, notable files, etc in the context. The is how Sourcegraph's Cody product works, for example.

Try deepseek 6.7B

It is a French company so maybe they have extra French datasets?

I've been quite disappointed by French LLMs on Huggingface when I tried a month ago.

I have the same impression. It is less verbose and doesn't beat around the bush (unlike GPT-3.5/4.5-Turbo), provides almost the same quality of code for the test cases I have, and has similar (GPT-4) or much better (GPT-3.5) spatial comprehension. It is at the GPT-3.5 level of math (read it as not good enough but better than anything else).

It is indicated by the magnet link which they have posted on twitter (see comments above).

I tried a bunch of my recent prompts to GPT-4 from daily use - this was often just slightly worse, sometimes slightly better. Fast too (tokens per second) while also not being overly wordy - very much appreciated that.

Refusals are a bit "I am just a language model"-y which GPT-4 has gotten away from. Also it's more refuse-y if I broach something rudely (which again I've found GPT-4 to have become much better at.)

Way better at everything than whichever Gemini I've been trying recently (can't tell for sure what I'm using when I use it.) But that one isn't even in contention for any use at all IME.

Overall it felt like I need to try it in daily use to work out if it's a contender with GPT-4 as a daily driver.

It doesn't on mine it still is really wrong but compared to mixtral when you tell next that it did a mistake it corrects it to the right answer where mixtral was making things up that were even more wrong.

Models don't "know" what they are. Usually the system prompt contains that imformation. If it doesn't, it will hallucinate an answer for you.