[HN Gopher] Mistral Saba
       ___________________________________________________________________
        
       Mistral Saba
        
       Author : stephen37
       Score  : 130 points
       Date   : 2025-02-17 13:56 UTC (9 hours ago)
        
 (HTM) web link (mistral.ai)
 (TXT) w3m dump (mistral.ai)
        
       | Cyph0n wrote:
       | Context on the name: https://en.wikipedia.org/wiki/Sheba
        
         | jorams wrote:
         | That makes a lot more sense than the Saba I initially thought
         | of[1]. The "specific geographies" seemed a bit overly specific.
         | 
         | [1]: https://en.wikipedia.org/wiki/Saba_(island)
        
       | yodon wrote:
       | > Mistral Saba is a 24B parameter model trained on meticulously
       | curated datasets from across the Middle East and South Asia.
        
       | elashri wrote:
       | That's interesting. It would be interesting to compare how this
       | will fare against Fanar (Arabic oriented models) [1]. I got
       | access to their API last week but still didn't play with it. I
       | think they did pretty good job with arabic dialects [2]. I don't
       | know if they have any plans to release weights though. There are
       | two models one trained from scratch and the other ia fine turned
       | of Google's Gemma.
       | 
       | Saba vs fanar. I like the names too.
       | 
       | [1] https://fanar.qa/en
       | 
       | [2] https://arxiv.org/abs/2501.13944
        
       | throwaway638637 wrote:
       | It says south asia but the blog post is about Arabic. Where are
       | the numbers on Urdu?
        
       | hazrmard wrote:
       | It's great to see proliferation of models in other languages!
       | 
       | Shoutout to Alif, a finetune of Llama 3 8b on Urdu datasets:
       | https://huggingface.co/large-traversaal/Alif-1.0-8B-Instruct
       | 
       | It'd be great to see a comparison.
        
       | laserduck wrote:
       | I wonder why they grouped languages from the Middle East and
       | South Asia together. Arabic and Hebrew are Semitic languages - no
       | language from that family tree is native to the subcontinent. It
       | would make sense if northern languages like Hindi, Urdu, Bengali,
       | Nepali, etc were grouped with Persian, French, Russian, etc since
       | those are all from the Indo-European family. South Indian
       | languages like Telugu and Tamil are from a completely different
       | family (Dravidian).
       | 
       | Why not either train the model exclusively on Semitic languages
       | for further performance for those languages or on a wider set of
       | languages for better multilingual performance overall? I don't
       | understand the logic here.
        
         | jazzyjackson wrote:
         | There is a lot of Indian laborers in the Middle East, so it's
         | not that Tamil and Arabic are related, but a model used for
         | that region should be fluent in both
        
           | reallymental wrote:
           | Not sure if you've been to the middle east, but there's no
           | way the labourers will have access to the internet besides
           | their phones. And those phones can only be used to
           | communicate with their loved ones back home, using Whatsapp.
           | 
           | They don't (or rather CAN) care about anything else in the
           | world.
           | 
           | They have a lot more problems than "this model doesn't
           | convert urdu to arabic well".
        
             | viraptor wrote:
             | WhatsApp is a gateway to many things. Meta's Ai is on
             | WhatsApp. Openai is on WhatsApp
             | https://www.hindustantimes.com/technology/whatsapp-users-
             | can... and I would really expect this model to have a
             | WhatsApp gateway as well.
             | 
             | > They have a lot more problems than "this model doesn't
             | convert urdu to arabic well".
             | 
             | I get what you mean, but I'm not sure what point you're
             | trying to make. That they're a lost cause with too many
             | problems and we shouldn't care about that use case? Why
             | wouldn't we want to create models to provide more
             | capabilities / information regardless?
        
             | slim wrote:
             | they still need to communicate with local administration
             | via "citizen app"(whatever it is called) to access any
             | service, pay their fines, etc...(and be tracked) I'm
             | guessing the stake holder in this project is the government
             | of qatar
        
         | sangnoir wrote:
         | > I wonder why they grouped languages from the Middle East and
         | South Asia together
         | 
         | Geography
        
         | achrono wrote:
         | There are a far greater number of speakers of Arabic in Germany
         | (1.4M) [1] than in Afghanistan (420K) [2].
         | 
         | So properly speaking, they should be advertising the target
         | region as Europe, Middle East and Africa. [3]
         | 
         | [1] https://en.wikipedia.org/wiki/Languages_of_Germany [2]
         | https://en.wikipedia.org/wiki/Languages_of_Afghanistan [3]
         | https://en.wikipedia.org/wiki/List_of_countries_and_territor...
        
         | viraptor wrote:
         | There's going to be diminishing returns in splitting the
         | languages where you get less information related to the region
         | / concept just because you're avoiding mixing languages. The
         | language was not the only aspect: "cultural background, and in-
         | depth regional knowledge". There's going to be lots of
         | information shared in south/North languages just because of the
         | geographically close (relatively anyway) distance.
         | 
         | I mean you wouldn't want to split a model into 3 separate ones,
         | where one contains Austrian, another Slovakian, and another
         | Hungarian, since there's going to be lots of cultural overlap.
        
       | diggan wrote:
       | Considering they don't talk about licensing, one can assume this
       | is proprietary?
       | 
       | ~2 years ago (Sep 27, 2023), Mistral AI said:
       | 
       | > we believe that an open approach to generative AI is necessary.
       | Community-backed model development is the surest path to fight
       | censorship and bias in a technology shaping our future. We
       | strongly believe that by training our own models, releasing them
       | openly, and fostering community contributions, we can build a
       | credible alternative to the emerging AI oligopoly. Open-weight
       | generative models will play a pivotal role in the upcoming AI
       | revolution.
       | 
       | > Mistral AI's mission is to spearhead the revolution of open
       | models.
       | 
       | https://mistral.ai/en/news/about-mistral-ai
       | 
       | Did something change since then, or why did they have a change of
       | hearts? Are they just doing a "OpenAI" and appear to believe in
       | something in order to further their own cause, or does it have
       | some particular reason behind it?
        
         | JimDabell wrote:
         | > One of the many custom-trained models to serve specific
         | geographies, markets, and customers
         | 
         | > Mistral Saba is a result of working closely with strategic
         | regional customers to address very specific challenges in
         | addressing bespoke use cases.
         | 
         | It seems like a customer paid them to train this model, so
         | presumably that customer gets to decide on licensing terms.
         | 
         | Isn't this Mistral's business model? Make general purpose
         | models available as open-source and train more specific models
         | for their customers?
        
           | diggan wrote:
           | It's available on their API service, so I'd assume it's not
           | one of the private models others pay them to create,
           | otherwise it would be fully private to the customer. Or, the
           | customer wants it to be released/open, then the customer
           | would release it, not Mistral. But I might misunderstand how
           | they operate, happy to correct what I understand.
           | 
           | Edit: Actually, it is outlined in the bottom of the post:
           | 
           | > we have also begun to train models for strategic customers
           | with the power of their deep and proprietary enterprise
           | context. These models stay exclusive and private to the
           | respective customers. If you would like to explore custom
           | training with Mistral AI, explore our applied AI offerings,
           | or please contact us.
           | 
           | So this is not one of those, as then it would be exclusive
           | and private to the customer. This (Saba) is one of the models
           | that I understood they would have released as at least "open-
           | weights", if following their initial goals according to the
           | early blog posts.
        
       | Terretta wrote:
       | GPT-4o mini keeps quietly demonstrating value per cost.
        
         | homarp wrote:
         | care to elaborate ?
         | 
         | Saba: input $0.2/M tokens / Output $0.6 /M tokens
         | 
         | GPT-4o Input: $0.15/M tokens Cached input:$0.075/M tokens
         | Output:$0.6/1M tokens
         | 
         | sources: https://openai.com/api/pricing/ and
         | https://mistral.ai/en/products/la-plateforme#pricing
        
       ___________________________________________________________________
       (page generated 2025-02-17 23:02 UTC)