Ajay retweeted
inkling 2 mini is the first small open model I'd actually try for real coding work on my own machine
Ajay retweeted
Today we’re releasing Inkling 2 Mini, a looped transformer, with open weights.
A big reason we're doing this is to let the AI safety community study looping, especially how loops affect chain-of-thought monitorability and how to interpret the computation inside loops. 🧵
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
Ajay retweeted
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
This is how @SarvamAI employees speak typically
I made a post on their founders calling India “data colonies of the west” btw when their own employees use the same foreign tools
They just troll but hardly answer any questions
1
171
Ajay retweeted
Replying to @rivuchakraborty @ajaysagar99
No need to google randos online. Nobody gives a fk what you're "building" unless you can take some criticism on your products.
Ajay retweeted
We’re looking for a founder to build and lead Notion India. DM me if you’re interested or know someone exceptional.
Replying to @karansingh010
You missed Step 0.
Like everything else in #DigitalIndia Sarvam was always a pure sarkari contracts play with Billionaire tech-idiot @NandanNilekani saar's blessings and all the PR is only to justify regulatory capture in the name of "sovereign AI".
xitter.gq/kingslyj/status/202429…
Replying to @kingslyj @NandanNilekani
You and I can't even bid for govt contracts without years of proven experience/turnover/profitability.
Rules don't apply for lapdogs.
July 2023 - @SarvamAI /Axonwise Private Limited incorporated.(August acc. PR)
March 2025 - @UIDAI contract awarded.
indianstartupnews.com/news/u…
Ajay retweeted
Replying to @logic
There is no secret to inference. DeepSeek has open sourced its inference pipeline. I'm sure many startups would compete with Sarvam on inference pricing. This is the race to the bottom.
Ajay retweeted
this was an entertaining rabbit hole
ankitg.me/blog/vibecoding-in…
Honestly I didn’t expect my post to blow up, I really do admire what smallest ai, maya, gnani do.
I just use twitter to be aware of new ai model releases
I work a bit on language models which gave me some background.
Now back to my usual job, building and my regular life!
1
1
373
Ajay retweeted
India’s real problem is trying to build a cheaper copy-paste economy and calling it innovation.
Ajay retweeted
Replying to @itsOmSarraf_ @ajaysagar99
No hatred but isnt sarvam running on subsidized compute from taxpayer money and the institution its running is also by govt and the funding that has been raised from vc which run on aif cat 2 funds which are pension funds of general public
So the so called token richness you are trying to prove is essenstially an epitome of true aristocracy behaviour by enslaving the tax payers ?
Le sarvam engineer on twitter-
guys we are building the frontier guys support us
Le sarvam founder - we are data colony of the west
Internally - uses Claude code, doesn’t uses his company’s own coding agents
A big W for Anthropic, L for state of Indian ai
Ajay retweeted
The problem i have in Sarvam being inference provider - is to centralised inference infra.
They are defacto Kaku (National) inference corporation of Aadhaaristan.
If GoI wanted to give GPU / boost AI economy. Decentralise to districts and make the growth distributed, equitable
I have no problem with Sarvam serving Chinese models or launching a provider. Here's why:
1. Sarvam needs to build its monetization and for that they have to compete in the open market. The primary PMF for models is coding. They're targeting coders and that's great. The only thing I'd say is that the I checked the pricing and apart from access to Sarvam models, the DeepSeek-v4.1-flash is the same as market pricing on Opencode and command code. There's very little reason to switch to Token factory.
2. You have to check what the Indian government got contractually for giving Sarvam discounted GPUs, and whether delivery was more than just announcements at the India AI summit. They're a business, and they have to work in shareholder interest. Nationalism is a pitch that got them the GPUs. They're not expected to deliver anything beyond their contract. Also, launching token factory doesn't mean they'll stop working on a foundational model.
3. If they're struggling to get usage of their models, and if they're training on usage via token factory that this is a good move to try and foster usage of their models for training.
4. We need to understand what's going on wrt models: for most use cases, models are commoditised: they're getting better faster but for most use cases, three-six months ago was good enough.
China is open sourcing and commoditising its models because its competitive advantage lies in hardware and access to rare earth metals. That's why the US govt is interested in Greenland and why US companies are lobbying the US govt against open source and open weights models: they don't want the model layer commoditised.
I have no issues with Sarvam serving Chinese models and hopefully fine tuning them based on Indian usage because other providers are already doing that anyway.
Someone said to me that India is consuming 20% of the worlds compute. We should use Chinese models and serve them locally to hopefully reduce the foreign exchange outflow.
Ajay retweeted
Replying to @vishnuvig
Incompetency and Impotency is much bigger problem than hosting Chinese AI models. This clearly shows they have diluted the vision and never had the substance to challenge the big techs.
And worst of all they lobbied, blocked other small companies from getting to India AI mission
Ajay retweeted
The "Sovereign Indian AI" playbook:
1. Get picked by the govt as India's chosen frontier lab
2. Collect GPU subsidies (sponsored by taxpayers)
3. Raise Millions (234) at a Billion valuation (1.5)
4. Launch a LLM inference service (Token Factory) that mostly serves Chinese SOTA model DeepSeek, Kimi, GLM, and Qwen.
5. Make sure to charge in INR with GST billing
Congratulations this is "Atmanirbhar AI" means using Indian GPUs to resell Chinese open models.
Peak ecosystem favoritism. If we wanted an OpenRouter clone, a three-person team in Bengaluru or Bhopal could have built it without a $1.5B valuation.
Made in 🇨🇳 , Served in 🇮🇳
















![(identity '[:pankaj :λ]) (@pankajdoharey) profile picture](https://pbs.twimg.com/profile_images/1748379845397184512/BXfj6M6z_400x400.jpg)












