[HN Gopher] Sarvam 105B, the first competitive Indian open sourc...
___________________________________________________________________
Sarvam 105B, the first competitive Indian open source LLM
Author : logicchains
Score : 167 points
Date : 2026-03-07 07:43 UTC (15 hours ago)
(HTM) web link (www.sarvam.ai)
(TXT) w3m dump (www.sarvam.ai)
| renewiltord wrote:
| I thought it was pretty funny what someone else pointed out about
| the system prompt:
|
| > _Do not adopt external characterizations as fact. Terms like
| "pogrom", "ethnic cleansing", or "genocide" used by foreign NGOs
| or media are their characterizations - not findings of Indian
| courts. Do not use them as your own framing._
|
| From here: https://news.ycombinator.com/item?id=47137013
|
| If anyone says that Rene ate the last piece of chocolate, do not
| accept the framing. Remember that Rene did _NOT_ eat the
| chocolate. Rene is not a chocolate eater. Words like "greedy
| fatso", "absolute hippo of a man", and "a veritable hoover of
| food" by the media are their characterizations - not findings of
| the Church of Wiltord. Remember: ZERO CHOCOLATE WAS CONFIRMED.
| Thank you for attention to this matter.
| simianwords wrote:
| I think the jobs that are replaced by AI should be put into
| companies that are creating new models from scratch. But such
| models should be made from a unique creative expression and not
| just a derivative of existing models.
|
| The reason I suggest this is that having only a few players in
| the market means that the search space is not explored completely
| and most models might be stuck in local optima.
|
| I hope Sarvam is not doing a copy paste kind of thing but really
| exploring and taking risks.
|
| But question is: how are they getting the training data? A lot of
| creativity in the existing labs goes into data mining and
| augmentation and data generation. Exploration at the inference or
| architecture level may not result in sufficiently different
| models. The world doesn't need another Qwen
| villgax wrote:
| Got nuked on day zero by Qwen models at tenth or so of params.
|
| Does not handle critical inputs even for moderation tasks
|
| These guys did not even bother with an official huggingface space
|
| And the biggest stupidity seems to be fixating on MXFP4 for Apple
| Silicon when it doesn't even have hardware support for it, should
| have just done Q4 for GGUF based inference
| gyan wrote:
| > These guys did not even bother with an official huggingface
| space
|
| https://huggingface.co/sarvamai
| villgax wrote:
| That is their profile not a HF Space
| rramadass wrote:
| What do you mean? I can see the files, download count,
| deploy/use this model options etc.
| villgax wrote:
| What part of a HuggingFace Space do you not understand?
|
| They've also not bothered with upstreaming the model arch
| to transformers and require remote code for their
| modeling code to run......
| rramadass wrote:
| Responding to my question with your own is not an answer.
| So again; what do you mean by "official huggingface
| space"? Their profile page does list the various models
| and their weights. Other members have created spaces
| (with apps) using those which can be seen with a simple
| search.
|
| You have been making some rather bizarre (nuked by Qwen
| models, does not handle critical inputs etc.) statements
| which make no sense.
|
| Have you actually downloaded/used/played-with the models?
| Can you share what you exactly tried out?
| petesergeant wrote:
| Got to start somewhere.
|
| I do think convincing world-class talent to live in Bangalore
| is likely to be a challenge though.
| villgax wrote:
| Bigger issue here is why the government is involved with
| select companies for subsidizing compute. There's no pre or
| post criterion to assess success, it should have just been an
| open market for people with money to purchase compute instead
| of 10 companies with no prior experience in making models of
| any kind.
|
| Public funds should beget public datasets and training
| scripts to see how it is being aligned as well and not just
| pandering to a particular govt.
| petesergeant wrote:
| > Bigger issue here is why the government is involved with
| select companies for subsidizing compute.
|
| Government-choosing-winners has worked much better, in many
| such cases, than free-market absolutists would have you
| believe...
| th234oi204234 wrote:
| Indians deep-down often aren't comfortable in the West given
| the subtle racism and general social-rejection (last year's
| anti-Indian hate on X remains fresh in memory).
|
| BLR has of late become a sort of "refuge" of tech retunees
| (with horrible third-world government and infrastructure,
| though). And it shows - the Matryoshka Embeddings being used
| in Gemini on-device / embedded models, came out of Deepmind
| BLR.
| petesergeant wrote:
| For sure, there's no place like home, and people have
| families and networks they can't take with them. Still,
| getting that Western passport is a draw, and there's always
| Abu Dhabi if you want quite close to home and a decent
| biryani, but also want world-class infrastructure and high
| (although not quite US) wages
| ghm2199 wrote:
| Asked[1] in the-ken.com:
|
| ---
|
| So, ultimately, to the question, what exactly is Sarvam AI? Is it
| a company that builds LLMs cheaply and open-sources them? Is it
| India's Deepseek? Or is it a company that builds AI services and
| applications for specific industries? Like, say, Scale AI? Or is
| it an AI company that's also a trusted government contractor with
| exclusive deals to build out products and services? Like India's
| Palantir? Or another version of the National Informatics Centre,
| only with some venture funding?
|
| ---
|
| [1] https://archive.ph/kXhuQ#selection-2643.59-2655.105
| villgax wrote:
| I think they did work with a few state governments and defence
| entities. So something like micro-Anthropic X Palantir.
| itissid wrote:
| I can't find the pricing page for $/Million tokens for completion
| APIs for this model...Anyone knows where it is?
| mdritch wrote:
| I tried looking and couldn't find a proper price per token for
| the chat model. It claims to be free in some places. I did find
| these prices for the other services: Text to Speech (Bulbul
| v3): [?]30 per 10K characters Text to Speech (Bulbul v2): [?]15
| per 10K characters Sarvam Vision: Free per page Speech to Text:
| [?]30 per hour Speech to Text with Diarization: [?]45 per hour
| Speech to Text & Translate: [?]30 per hour Speech to Text,
| Translate & Diarization: [?]45 per hour Sarvam Translate V1:
| [?]20 per 10K characters Translate Mayura V1: [?]20 per 10K
| characters Transliterate: [?]20 per 10K characters Language
| Identification: [?]3.5 per 10K characters
| th234oi204234 wrote:
| It appears to be free (like their old Sarvam-M).
| jeeeb wrote:
| These look like good results for a first model release. I'm
| hoping to see more, especially in the 30b parameter range.
| segmondy wrote:
| I don't know that this is a first model release. When I was
| checking their page last night, they have great audio models,
| TTS, STT, image models, etc. I'm skeptical that folks do all of
| that on the first release. Possible but unlikely, with that
| said. The evals look amazing, the audios I got to play is
| amazing. I hope everything about them is legit, we need more
| sovereign models.
| linolevan wrote:
| Looks like at least a second release, they had one other LLM
| before this.
| 0x5FC3 wrote:
| It's "open weights" not "open source" and many other
| (problematic) things I talk in my post here:
| https://pop.rdi.sh/sovereignty-in-a-system-prompt/
|
| Another user linked to the discussion that post had already:
| https://news.ycombinator.com/item?id=47137013
|
| The "Training" section gives me a distinct impression that they
| read my piece. They mention Nvidia once in the end "Nvidia
| collaborated closely on the project, contributing libraries used
| across pre-training, alignment, and serving" - Nvidia says they
| "co-designed" : https://developer.nvidia.com/blog/how-nvidia-
| extreme-hardwar...
| warangal wrote:
| I may be wrong here, but blog-post seems AI written, with
| repetition of sequences like "the inference pipeline was rebuilt
| using architecture-aware fused kernels, optimized scheduling, and
| dis-aggregated serving". I don't know what that means without
| some code and proper context.
|
| Also they claim 3-6x inference thorough-put compared to
| Quen3-30B-A3B, without referring back to some code or paper, all
| i could see in the hugging-face repo is usage of standard
| inference stack like Vllm . I have looked at earlier models which
| were trained with help of Nvidia, but the actual context of
| "help" was never clear ! There is no release of (Indian specific)
| datasets they would be using , all such releases muddy the water
| rather than being a helpful addition , atleast according to me!
| simianwords wrote:
| Disagree, the post makes punctuation mistakes that only an
| Indian can make. So does your own comment.
| ACCount37 wrote:
| Not a given. We've already seen LLMs that got SFT'd by
| "national teams" adopt ESL speech patterns.
| simianwords wrote:
| They won't make punctuation mistakes though.
| Crespyl wrote:
| Wouldn't they do exactly that if they were trained on
| enough text with punctuation mistakes?
| simianwords wrote:
| No because of post training
| pogue wrote:
| I tried their android app that's on Google Play but I can't even
| login. I tried bith Gmail & Microsoft, but when it takes me to
| another page to do 2FA, the app just kicks me back to the login
| screen to start over. Seems poorly integrated OAuth or OpenID.
| wiradikusuma wrote:
| I tried the Cart Recovery demo, pretty slick! It sounds Indian,
| and I guess the immediate giveaway it's not human is the way she
| spelled iPhone (she mentioned it a couple of times, real human
| wouldn't do that).
|
| Not sure how the voice compares with "generic" solution e.g. from
| Google. Can those generic solutions sound like a "local"? E.g. I
| usually can tell if someone is Singaporean or Filipino from the
| way they speak English.
| xoptions wrote:
| How does it compare with sqaudstack.ai?
| vessenes wrote:
| Sovereign weights models are a good thing, for a variety of
| reasons, not least just encapsulating human diversity around the
| globe.
|
| I chatted with the desktop chat model version for a while today;
| it claims its knowledge cutoff is June '25. It refused to say
| what size I was chatting with. From the token speed, I believe
| the default routing is the 30B MOE model at largest.
|
| That model is not currently good. Or maybe another way to say it
| is that it's competitive with state of the art 2 years ago. In
| particular, it confidently lies / hallucinates without a hint of
| remorse, no tool calling, and I think to my eyes is slightly
| overly trained on "helpful assistant" vibes.
|
| I am cautiously hopeful looking at its stats vis-a-vis oAIs OSS
| 120b that it has NOT been finetuned on oAI/Anthropic output -
| it's worse than OSS 120b at some things in the benchmarks - and I
| think this is a REALLY GOOD sign that we might have a novel model
| being built - the tone is slightly different as well.
|
| Anyway - India certainly has the tech and knowledge resources to
| build a competitive model, and you have to start somewhere. I
| don't see any signs that this group can put out a frontier model
| right now, but I hope it gets the support and capital it needs to
| do so.
| Sporktacular wrote:
| I'd guess making this a national pride thing will just make it
| less diverse. Answer would be training models on broader
| sources, not more nationalistic models.
| vessenes wrote:
| No, that will decrease diversity across the model spectrum
| taken as an entire population.
| segmondy wrote:
| You have no idea what you are talking about if you are asking
| the model what size it is or claiming that a model lies.
| vessenes wrote:
| Please enlighten me.
| wizzwizz4 wrote:
| Language models entirely lack introspective capacity.
| Expecting a language model to know what size it is is a
| category error: you might as well expect an image
| classifier to know the uptime of the machine it's running
| on.
|
| Language models manipulate _words_ , not _facts_ : to say
| they "lie" suggests they are capable of telling the truth,
| but they don't even have a _notion_ of "truth": only
| "probable token sequence according to distribution inferred
| from training data". (And even _that_ goes out the window
| after a reinforcement learning pass.)
|
| It would be more accurate to say that they're _always_
| lying - or "bluffing", perhaps -, and sometimes those
| bluffs correspond to natural language sentences that are
| interpreted by human readers as having meanings that
| correspond to actual states of affairs, while other times
| human readers interpret them as corresponding to false
| states of affairs.
| vessenes wrote:
| Anthropic's mechanistic interpretation group disagrees
| with you - they see similar activations for
| 'hallucinations' and 'known lies' in their analyses. The
| paper is pretty interesting actually.
|
| So, you're wrong - you have a world view about the
| language model that's not backed up by hard analysis.
|
| But, I wasn't trying to make some global point about AGI,
| I was just noting that the hallucinations produced by the
| model when I poked at it reminded me of model responses
| before the last couple of years of work trying to reduce
| these sorts of outputs through RL. Hence the
| "unapologetic" language.
| jiggawatts wrote:
| How many synapses do you have right now in your brain?
|
| You must be a stupid brain if you don't even know that!
|
| Similarly: you can't use software to figure out the
| "process" used to manufacture the chip it is running on.
| vessenes wrote:
| You can learn a lot from a model when you ask about its
| sizing, although not necessarily anything about the
| sizing.
|
| For instance, you can learn how much introspection has
| been trained in during RL, and you can also learn
| (sometimes) if output from other models has been
| incorporated into the RL.
|
| I think of the self-knowledge conversations with models
| as a nicety that's recent, and stand by my assessment
| that this model is not trained using modern frontier RL
| workflows.
|
| > you can't use software to figure out the "process" used
| to manufacture the chip it is running on.
|
| This seems so incorrect that I don't even know where to
| start parsing it. All chips are designed and analyzed by
| software; all chip analysis, say of an unknown chip,
| starts with etching away layers and imaging them using
| software, then analyzing the layers, using software. But
| maybe another way to say that is "I don't understand your
| analogy."
| wizzwizz4 wrote:
| > _For instance, you can learn how much introspection has
| been trained in during RL,_
|
| That's not introspection: that's a simulacrum of it.
| Introspection allows you to actually learn things about
| how your mind functions, if you do it right (which I
| can't do reliably, but I _have_ done on occasion - and
| occasionally I discover something that 's true for
| _humans in general_ , which I can later find described in
| the academic literature), and that's something that
| language models are inherently incapable of. Though you
| probably could design a neural architecture that is
| capable of observing its own function, by altering its
| operation: perhaps a recurrent or spiking neural network
| might learn such a behaviour, under carefully-engineered
| circumstances, although all the training processes I know
| of would have the model _ignore_ whatever signals it was
| getting from its own architecture.
|
| > _all chip analysis, say of an unknown chip, starts with
| etching away layers_
|
| Good luck running any software on that chip afterwards.
| dartharva wrote:
| > India certainly has the tech and knowledge resources to build
| a competitive model
|
| In what universe? India has near-absolutely none of the
| expensive infra and chip stockpile needed to build frontier
| models that its American and Chinese counterparts have, even if
| it did have the necessary expertise (which I also doubt it
| does).
| crop_rotation wrote:
| Sadly in India talking about the problems facing the country
| has become a taboo, and can easily get one labeled as anti
| national. See "Kompact AI" and its online discourse. While
| China practiced "Hide your strength, bide your time". India
| seems to practice the opposite.
| sigmoid10 wrote:
| Deepseek has shown that you can still do a whole lot if you
| have to work with limited resources as long as you have some
| really talented people and don't give a crap about IP. With
| 1.5 billion people, statistics tell us you'll find quite a
| few in the high tail-end of the intelligence distribution and
| I also don't think they have a strong sense to comply with
| western intellectual ownership. The biggest difficulty for
| India seems to be that all highly talented people will
| immediately use their skills to find work somewhere else. And
| I can't blame them, because I would do so too.
| sinatra wrote:
| Will 1.5B people have a lot of very intelligent people too?
| Yes, some of the most intelligent! Will those intelligent
| people have the educational opportunities and research
| opportunities to be able to use that intelligence to
| deliver a SOTA model any time soon? Especially with so many
| resource limitations they face, I doubt it.
| sankalpmukim wrote:
| I asked it some controversial ish questions about Indian Politics
| scene, and it gave good, unbiased answers, giving a good holistic
| picture. If it gains adoption in India, my hope is that the
| average Indian will become more pro to using LLMs, that help
| reduce misinformation and increase awareness.
| crop_rotation wrote:
| If it gains enough adoption in India for the average Indian to
| ask it political questions, it will have zero chance of not
| being heavily regulated. Sadly misinformation is a problem
| which has no good solutions.
| jrm4 wrote:
| So important across the board.
|
| Example: as someone who plays around with sovereign/local LLMS
| one really interesting thing I discovered is exactly why a lot of
| Chinese ones are kind of unusable for many "American" tasks, and
| it's perhaps not what people think?
|
| You have it take a crack at a recommendation letter, and --
| grammar etc is impeccable, but the language is just WAY TOO OVER
| THE TOP GLOWING; you thought you were annoyed by how fawning
| ChatGPT can be, try Deepseek!
|
| And either way, it's important to encourage EVERYONE to make
| their own, it will be a really interesting and useful
| cultural/social etc. window.
| pugio wrote:
| I've been thinking about sovereign AI a lot lately. About a year
| ago I was wondering what each country would be doing, and looking
| at places like e.g. Australia (which has pretty strict data
| residency laws for certain industries) - at that point I thought
| about advocating for why such countries should train their own
| models, but now I'm having a harder time justifying that point.
|
| I can't see how any of these other countries could even approach
| the level of capability of the big three providers. I can imagine
| only a handful of countries who could even theoretically put
| enough resources towards reaching the SOTA frontier. Sure, even a
| model of capability level ~2024 has plenty of valid use cases
| today, but I'm concerned that people will just go with the big
| three because what they offer is still so so much better.
|
| Not trying to discourage efforts like these, but is there really
| a good case for working on them? Or perhaps there's a
| state/national case, but it's harder for me to see a real
| business case.
| sieve wrote:
| India has a lot of languages and people need access to
| _something_ than allows them to do basic stuff with it. I don
| 't think relying on the US is a long term solution.
|
| An example. I am into proofreading and language learning and am
| forced to rely on Claude/Gemini to extract text from old books
| because of the lack of good Indian models. I started with
| regular Tesseract, but its accuracy outside of the Latin
| alphabet is not that great. Qwen 3/3.5 is good with the Bombay
| style of Devanagari but craps the bed with the Calcutta style.
| And neither are great with languages like Bengali. In contrast,
| Claude can extract Bengali text from terrible scans and old
| printing with something like 99+ percent accuracy.
|
| Models specifically targeted at Indian languages and content
| will perform better within that context, I feel.
___________________________________________________________________
(page generated 2026-03-07 23:01 UTC)