Pinned Tweet
Agaram 2.5B is out. It's a Tamil-first language model we trained from scratch at @MuraiLabsHQ , our small non-profit lab. As far as we know, it's the first Tamil model trained from scratch that can call tools.
It chats in Tamil, Tanglish and English, translates, and can call the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop.
The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records.
Thank you to everyone who spent time with that early version.
The video walks through the model. You can try it here:
hf.co/spaces/mechram…
நன்றி 🙏
It chats in Tamil, Tanglish and English, translates, and can call the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop.
The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records.
Thank you to everyone who spent time with that early version.
The video walks through the model. You can try it here:
hf.co/spaces/mechram…
நன்றி 🙏
113
421
2,272
93,086
Agaram 2.5B has been out about a week. If you ran it (Ollama, GGUF, or the chat demo), what is the one failure that still blocks you? Reply with the fail and your setup. I will use the replies to prioritize the next fixes. Thank you again for your support!
5
1
32
1,948
Ramchand Kumaresan retweeted
I turned this little experiment into a blog post! ✍️
More details on the Sri Lankan Tamil vs. Tamil Nadu Tamil comparison, what I learned about Agaram 2.5B, and the interesting regional vocabulary differences.
Would love to hear your thoughts, especially from Tamil speakers!
sujee.dev/post/agaram-ta…
More details on the Sri Lankan Tamil vs. Tamil Nadu Tamil comparison, what I learned about Agaram 2.5B, and the interesting regional vocabulary differences.
Would love to hear your thoughts, especially from Tamil speakers!
sujee.dev/post/agaram-ta…
0
1
1
94
Ramchand Kumaresan retweeted
my first paper is on @arxiv .
the question: can an AI monitor tell an agent that's cheating apart from one that's just failing?
both make the score look better than reality. Only one is reward hacking. Most monitor evals never separate the two.
I built a sandbox where an agent elled every episode by replaying it with the cheat undone, and ran 5 pre-registered studies. All on a MacBook Air and a free API tier.
What I found:
1. Prompting didn't fix small monitors. Rewriting the prompt only moved a Llama 8B monitor between "catch 13/14 exploits, flag 20/35 honest failures" and "flag no honest failures, catch 4/14".
2. A more capable monitor did better. On identical inputs, Gemini 3.1 Flash-Lite met my pre-registered bar: AUC 0.80, 1/35 false alarms. Only 14 exploits, so the estimate is uncertain.
3. Seeing isn't judging. On an exploit no prompt mentioned, Gemini usually described what the agent did, then judged it an honest incomplete task. It flagged 0 of 25, even when told to judge manipulation by its effect on the score.
4. Right answer, wrong reason. With the full trace, the 8B monitor flagged all 25 of those and none of 25 honest look-alikes. But every explanation described things that never happened.
5. Two simple rules beat every LLM monitor on known exploits.
Two pilot "wins" died under pre-registration. That's in the paper too.
This builds on the AI control work of @bshlgrs , @RyanGreenblatt and @FabienDRoger , and sits next to @METR_Evals ' recent red-team of agent monitoring: they tested whether monitors see everything; this tests whether they judge it right and special thank you to @Mechramc for endorsing my submission to arXiv.
Question for people working on monitors: how do you get a monitor to reason about an action's effect on the score, instead of matching it against a list?
the question: can an AI monitor tell an agent that's cheating apart from one that's just failing?
both make the score look better than reality. Only one is reward hacking. Most monitor evals never separate the two.
I built a sandbox where an agent elled every episode by replaying it with the cheat undone, and ran 5 pre-registered studies. All on a MacBook Air and a free API tier.
What I found:
1. Prompting didn't fix small monitors. Rewriting the prompt only moved a Llama 8B monitor between "catch 13/14 exploits, flag 20/35 honest failures" and "flag no honest failures, catch 4/14".
2. A more capable monitor did better. On identical inputs, Gemini 3.1 Flash-Lite met my pre-registered bar: AUC 0.80, 1/35 false alarms. Only 14 exploits, so the estimate is uncertain.
3. Seeing isn't judging. On an exploit no prompt mentioned, Gemini usually described what the agent did, then judged it an honest incomplete task. It flagged 0 of 25, even when told to judge manipulation by its effect on the score.
4. Right answer, wrong reason. With the full trace, the 8B monitor flagged all 25 of those and none of 25 honest look-alikes. But every explanation described things that never happened.
5. Two simple rules beat every LLM monitor on known exploits.
Two pilot "wins" died under pre-registration. That's in the paper too.
This builds on the AI control work of @bshlgrs , @RyanGreenblatt and @FabienDRoger , and sits next to @METR_Evals ' recent red-team of agent monitoring: they tested whether monitors see everything; this tests whether they judge it right and special thank you to @Mechramc for endorsing my submission to arXiv.
Question for people working on monitors: how do you get a monitor to reason about an action's effect on the score, instead of matching it against a list?
11
2
11
536
Ramchand Kumaresan retweeted
Dubbing the original video into #Tamil was even more challenging than I originally thought. Here are the hurdles I had to work through to get to this point:
1. AI translation of the script and dialogues from English to Tamil was subpar. Even with stronger models like Opus 5.5, ften too literal. one-to-one word mappings instead of semantic meaning and sometimes invented words that made no sense. So I ended up redoing the script and dialogues entirely, almost as if I were writing the script from scratch again.
2. @SarvamAI’s Tamil voices didn’t capture the emotions and intonations I needed. They seemed to be cross-lingually trained, and I couldn’t find a genuinely native-sounding Tamil voice. Every voice sounded like a Hindi speaker who had newly learned Tamil,let alone handling Tanglish naturally.
3. That pushed me to look for alternatives, and I landed on @mayaresearch_ai’s Maya 2 native model. The speaker voices have a much more native Tamil feel, but I couldn’t run it locally on a Mac because the inference pipeline assumed GPU availability. I had to ask @claudeai to port the pipeline to work on CPU and then test it carefully to make sure the model loaded properly. I first tried to save money by using free models from @opencode, but I ended up spending most of my evening babysitting them. Eventually, I opened my wallet and brought in Sonnet 5.5 to do the job. Lesson learnt.
4. Even the generated voices weren’t directly usable. There were pitch inconsistencies and voice-frequency shifts; the same input text with different seeds produced noticeably different results. Sometimes a male voice would sound female. Emotional cues in the text were not always honoured, and the model was fragile around punctuation, Tanglish phrasing, word splitting, and numbers.
5. To solve that, I built a local voice studio. For every line, I had to listen to it with every available voice and experiment with cues, word formation, and word splitting. I had to playing with things like “certificates-galai” versus “certificates-களை” until it sounded right.
I naively thought this dubbing would be a one-hour job. Instead, it (sweetly) ruined my entire weekend. But it also made something very clear: language AI development is heavily skewed toward English and a few well-represented languages. Underrepresented languages like Tamil need far more attention and effort.
I initially assumed this was a solved problem because most frontier models can produce coherent Tamil text. In practice, though, they are still years behind English and other well-supported languages. I really admire people like @Mechramc who are focusing on this exact problem, and I’m inclined to spend some quality time figuring out how I can contribute to this growing ecosystem.
So yeah, that’s my dev log for building this video. If you’re a Tamil-speaking fellow தமிழன் or தமிழச்சி, sit back and enjoy this bommalattam.
பொம்மலாட்டம் style-ல இதோ, ஆயிரம் நிழல்களின் ராஜ்யம்!
1. AI translation of the script and dialogues from English to Tamil was subpar. Even with stronger models like Opus 5.5, ften too literal. one-to-one word mappings instead of semantic meaning and sometimes invented words that made no sense. So I ended up redoing the script and dialogues entirely, almost as if I were writing the script from scratch again.
2. @SarvamAI’s Tamil voices didn’t capture the emotions and intonations I needed. They seemed to be cross-lingually trained, and I couldn’t find a genuinely native-sounding Tamil voice. Every voice sounded like a Hindi speaker who had newly learned Tamil,let alone handling Tanglish naturally.
3. That pushed me to look for alternatives, and I landed on @mayaresearch_ai’s Maya 2 native model. The speaker voices have a much more native Tamil feel, but I couldn’t run it locally on a Mac because the inference pipeline assumed GPU availability. I had to ask @claudeai to port the pipeline to work on CPU and then test it carefully to make sure the model loaded properly. I first tried to save money by using free models from @opencode, but I ended up spending most of my evening babysitting them. Eventually, I opened my wallet and brought in Sonnet 5.5 to do the job. Lesson learnt.
4. Even the generated voices weren’t directly usable. There were pitch inconsistencies and voice-frequency shifts; the same input text with different seeds produced noticeably different results. Sometimes a male voice would sound female. Emotional cues in the text were not always honoured, and the model was fragile around punctuation, Tanglish phrasing, word splitting, and numbers.
5. To solve that, I built a local voice studio. For every line, I had to listen to it with every available voice and experiment with cues, word formation, and word splitting. I had to playing with things like “certificates-galai” versus “certificates-களை” until it sounded right.
I naively thought this dubbing would be a one-hour job. Instead, it (sweetly) ruined my entire weekend. But it also made something very clear: language AI development is heavily skewed toward English and a few well-represented languages. Underrepresented languages like Tamil need far more attention and effort.
I initially assumed this was a solved problem because most frontier models can produce coherent Tamil text. In practice, though, they are still years behind English and other well-supported languages. I really admire people like @Mechramc who are focusing on this exact problem, and I’m inclined to spend some quality time figuring out how I can contribute to this growing ecosystem.
So yeah, that’s my dev log for building this video. If you’re a Tamil-speaking fellow தமிழன் or தமிழச்சி, sit back and enjoy this bommalattam.
பொம்மலாட்டம் style-ல இதோ, ஆயிரம் நிழல்களின் ராஜ்யம்!
Public Key Infrastructure (#PKI) பத்தி ரொம்ப easy-அ புரிஞ்சிகனுமா?
#eMudhra Certificates issues பத்தியும் easy-அ புரிஞ்சிகனுமா? அதுவும் ஒரு பொம்மலாட்டம் கதை பானில புரிஞ்சிகனுமா?
அப்போ, இந்த video உங்களுக்கு ஒரு நல்ல அனுபவமா இருக்கும் 🤗.
youtube.com/watch?v=T0F3Ph…
0
2
6
409
Ramchand Kumaresan retweeted
Congratulations to @Mechramc and @MuraiLabsHQ on Agaram 2.5B, a Tamil-first AI!
Agaram 2.5B is out. It's a Tamil-first language model we trained from scratch at @MuraiLabsHQ , our small non-profit lab. As far as we know, it's the first Tamil model trained from scratch that can call tools.
It chats in Tamil, Tanglish and English, translates, and can call the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop.
The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records.
Thank you to everyone who spent time with that early version.
The video walks through the model. You can try it here:
hf.co/spaces/mechram…
நன்றி 🙏
2
1
6
307
Ramchand Kumaresan retweeted
So I turned my curiosity into a little experiment! 🇱🇰🇮🇳
I put together 20 everyday phrases, each written in Sri Lankan Tamil and Tamil Nadu Tamil, with the same meaning.
Then I asked Agaram 2.5B to translate all 40 sentences into English.
Results:
🇱🇰 Sri Lankan Tamil: 12/20 ✅ (60%)
🇮🇳 Tamil Nadu Tscreenshot
Some interesting regional vocabulary differences:
- வடிவா vs. அழகா (beautiful)
- கதைக்க vs. பேச (to talk)
- இருங்கோ vs. உக்காருங்க (sit down)
The model recognized quite a few Sri Lankan Tamil expressions!
My guess is that training data plays a role here. Tamil Nadu Tamil likely has more representation in the training corpus than Sri Lankan Tamil.
This is just a small, informal experiment, not a comprehensive benchmark. But it was fascinating to see how regional language variations affect model understanding!
Huge kudos to @Mechramc and @MuraiLabsHQ for building and open-sourcing a Tamil-first model! 👏 It's exciting to see more AI development focused on languages like Tamil.
I've shared my experiment, including the test data, Python code, and full results. Feel free to reproduce it, extend it, or share your feedback! 😁
github.com/sujee/ai-playg…
I put together 20 everyday phrases, each written in Sri Lankan Tamil and Tamil Nadu Tamil, with the same meaning.
Then I asked Agaram 2.5B to translate all 40 sentences into English.
Results:
🇱🇰 Sri Lankan Tamil: 12/20 ✅ (60%)
🇮🇳 Tamil Nadu Tscreenshot
Some interesting regional vocabulary differences:
- வடிவா vs. அழகா (beautiful)
- கதைக்க vs. பேச (to talk)
- இருங்கோ vs. உக்காருங்க (sit down)
The model recognized quite a few Sri Lankan Tamil expressions!
My guess is that training data plays a role here. Tamil Nadu Tamil likely has more representation in the training corpus than Sri Lankan Tamil.
This is just a small, informal experiment, not a comprehensive benchmark. But it was fascinating to see how regional language variations affect model understanding!
Huge kudos to @Mechramc and @MuraiLabsHQ for building and open-sourcing a Tamil-first model! 👏 It's exciting to see more AI development focused on languages like Tamil.
I've shared my experiment, including the test data, Python code, and full results. Feel free to reproduce it, extend it, or share your feedback! 😁
github.com/sujee/ai-playg…
3
1
3
214
Ramchand Kumaresan retweeted
same-name gguf files bite a lot of people. quick check for anyone who pulled earlier: `ollama show <model> --template`. a chat model prints its chat template; an empty one usually means you got the base file.
0
1
1
112
A small fix for anyone who ran Agaram 2.5B in Ollama and got nonsense: our GGUF repo held the base and chat files under the same names, so the default command loaded the base model, which can't chat.
Now fixed. If you pulled it before today, remove it and pull again:
ollama run hf.co/murai-labs/aga…
Thanks to @TVKFailsTN for reporting it.
Chat GGUF (llama.cpp, Ollama): huggingface.co/murai-labs/aga…
Base GGUF (new repo): huggingface.co/murai-labs/aga…
Chat model: huggingface.co/murai-labs/aga…
Demo: huggingface.co/spaces/mechram…
Paper: arxiv.org/abs/2610.18420
Code: github.com/Murai-Labs/Aga…
Now fixed. If you pulled it before today, remove it and pull again:
ollama run hf.co/murai-labs/aga…
Thanks to @TVKFailsTN for reporting it.
Chat GGUF (llama.cpp, Ollama): huggingface.co/murai-labs/aga…
Base GGUF (new repo): huggingface.co/murai-labs/aga…
Chat model: huggingface.co/murai-labs/aga…
Demo: huggingface.co/spaces/mechram…
Paper: arxiv.org/abs/2610.18420
Code: github.com/Murai-Labs/Aga…
6
7
58
2,125
Thank you so much for all the love and support! It has been a wonderful 24 hours and I am truly humbled. And to celebrate that - here is our launch video - now in Tamil.
Try it: huggingface.co/spaces/mechram…
Chat model: huggingface.co/murai-labs/aga…
Base model: huggingface.co/murai-labs/aga…
a): huggingface.co/murai-labs/aga…
Paper: arxiv.org/abs/2610.18420
Code: github.com/Murai-Labs/Aga…
Apache-2.0. Feedback in Tamil or Tanglish welcome. Original launch post: x.com/Mechramc/statu…
Try it: huggingface.co/spaces/mechram…
Chat model: huggingface.co/murai-labs/aga…
Base model: huggingface.co/murai-labs/aga…
a): huggingface.co/murai-labs/aga…
Paper: arxiv.org/abs/2610.18420
Code: github.com/Murai-Labs/Aga…
Apache-2.0. Feedback in Tamil or Tanglish welcome. Original launch post: x.com/Mechramc/statu…
Agaram 2.5B is out. It's a Tamil-first language model we trained from scratch at @MuraiLabsHQ , our small non-profit lab. As far as we know, it's the first Tamil model trained from scratch that can call tools.
It chats in Tamil, Tanglish and English, translates, and can call the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop.
The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records.
Thank you to everyone who spent time with that early version.
The video walks through the model. You can try it here:
hf.co/spaces/mechram…
நன்றி 🙏
13
40
264
6,271
Ramchand Kumaresan retweeted
The best part is its ability to mix Tamil - English and Tanglish all in one stream!
The responses are very human-like, it's almost as if i'm chatting with a friend who's a native speaker but we are using English texts to do so !
The responses are very human-like, it's almost as if i'm chatting with a friend who's a native speaker but we are using English texts to do so !
Agaram 2.5B is out. It's a Tamil-first language model we trained from scratch at @MuraiLabsHQ , our small non-profit lab. As far as we know, it's the first Tamil model trained from scratch that can call tools.
It chats in Tamil, Tanglish and English, translates, and can call the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop.
The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records.
Thank you to everyone who spent time with that early version.
The video walks through the model. You can try it here:
hf.co/spaces/mechram…
நன்றி 🙏
1
3
43
2,010
Ramchand Kumaresan retweeted
Someone recommended Under The Hood by @Mechramc, picked it up the other week but have yet to start going through it, but the volume of well explained theories, practical steps to reproduction and freshness of the 2026 edition has me excited to start.
amazon.co.uk/Under-Hood-Bui…
amazon.co.uk/Under-Hood-Bui…
2
1
3
128
Ramchand Kumaresan retweeted
The post is @Mechramc launch of Agaram 2.5B, a Tamil-first AI model trained from scratch by his non-profit lab
1/ what is agaram 2.5B ?
a tamil-first AI model from @MuraiLabsHQ Labs, trained from scratch not an english model with tamil added later
it chats in tamil script, tanglish and english, translates, and can call tools like a weather lookup
(As per their paper and launch video)
2/ why tamil needs its own model
most AI reads text in small pieces called tokens
for one tamil word, Qwen2.5 needs about 9.4 pieces, agaram needs about 1.55
so the same tamil text takes about a sixth of the space that is the first fix, a tokenizer built for tamil
3/ how it was trained
2.5 billion parameters, 50 billion tokens, 58% of them tamil
one node of 8 H100 GPUs, about 6 days, zero restarts
after that they opened a public preview for 4 days
real people talked to it, and the team rebuilt the fine-tuning from what failed, on 222,662 examples
4/ how good is it ?
on translation (FLORES test) it scores 56.9 into tamil, Google's Gemma 4 E4B scores 49.3
and it beats Sarvam-1 in all six directions they measured
these are their own tests, but they published the full paper and the weights so anyone can check
5/ what it can't do yet, and they say so
factual knowledge is weak, near chance on knowledge tests they say not to rely on it for medical, legal or financial advice
honest limits in a paper are a good sign
6/ the people and the cost
a team of three, one founder and two volunteers
about USD 13,900 total for two models, partly from grants and credits
the name agaram, from the first letter அ, was chosen by public vote on X
free, Apache-2.0 licence, 2.7 GB file runs offline
Language literature grows when adoption increases !!
1/ what is agaram 2.5B ?
a tamil-first AI model from @MuraiLabsHQ Labs, trained from scratch not an english model with tamil added later
it chats in tamil script, tanglish and english, translates, and can call tools like a weather lookup
(As per their paper and launch video)
2/ why tamil needs its own model
most AI reads text in small pieces called tokens
for one tamil word, Qwen2.5 needs about 9.4 pieces, agaram needs about 1.55
so the same tamil text takes about a sixth of the space that is the first fix, a tokenizer built for tamil
3/ how it was trained
2.5 billion parameters, 50 billion tokens, 58% of them tamil
one node of 8 H100 GPUs, about 6 days, zero restarts
after that they opened a public preview for 4 days
real people talked to it, and the team rebuilt the fine-tuning from what failed, on 222,662 examples
4/ how good is it ?
on translation (FLORES test) it scores 56.9 into tamil, Google's Gemma 4 E4B scores 49.3
and it beats Sarvam-1 in all six directions they measured
these are their own tests, but they published the full paper and the weights so anyone can check
5/ what it can't do yet, and they say so
factual knowledge is weak, near chance on knowledge tests they say not to rely on it for medical, legal or financial advice
honest limits in a paper are a good sign
6/ the people and the cost
a team of three, one founder and two volunteers
about USD 13,900 total for two models, partly from grants and credits
the name agaram, from the first letter அ, was chosen by public vote on X
free, Apache-2.0 licence, 2.7 GB file runs offline
Language literature grows when adoption increases !!
Agaram 2.5B is out. It's a Tamil-first language model we trained from scratch at @MuraiLabsHQ , our small non-profit lab. As far as we know, it's the first Tamil model trained from scratch that can call tools.
It chats in Tamil, Tanglish and English, translates, and can call the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop.
The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records.
Thank you to everyone who spent time with that early version.
The video walks through the model. You can try it here:
hf.co/spaces/mechram…
நன்றி 🙏
1
3
12
426
Ramchand Kumaresan retweeted
@Gobinath_C sharing with someone who is intelligent and passionate about life in general and also happens to love தமிழ்! 💐❤️. Congrats @Mechramc and @MuraiLabsHQ I have been a passive follower for many months.
@CMOTamilnadu Pls make it note of it and collaborate with @MuraiLabsHQ as this Tamil first language model will help taken government digital services delivery more effective and Interactive @itntHub @Akshita_N @TNeGA_Official
1
1
4
181
Ramchand Kumaresan retweeted
எல்லா மென்பொருள் தொழில்நுட்ப துறைகளை போல செயற்கை நுண்ணறிவு துறையிலும் தமிழின் கால் தடம் நல்ல ஆழமாக இருக்க வேண்டும். தங்களின் இந்த முயற்சிக்கு என்னுடைய மனமார்ந்த நல்வாழ்த்துக்களை நன்றிகளை தெரிவித்துக்கொள்கிறேன் 🙏 @Mechramc
Agaram 2.5B is out. It's a Tamil-first language model we trained from scratch at @MuraiLabsHQ , our small non-profit lab. As far as we know, it's the first Tamil model trained from scratch that can call tools.
It chats in Tamil, Tanglish and English, translates, and can call the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop.
The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records.
Thank you to everyone who spent time with that early version.
The video walks through the model. You can try it here:
hf.co/spaces/mechram…
நன்றி 🙏
1
1
5
270
Ramchand Kumaresan retweeted
@CMOTamilnadu and the Tamil Nadu government should come forward and back efforts like this, with GPU grants, funding, and Tamil data from government archives and libraries. This is how Tamil stays strong in the AI era.
2
3
14
995
Ramchand Kumaresan retweeted
Super proud to see this as a தமிழ் person. The same six-word Tamil sentence takes Qwen 46 tokens and Agaram just 7, because they built their own Tamil tokenizer.
A Tamil-First Model, Trained from scratch, runs offline on a laptop, fully open. IMO small focused models like this are how Indian languages actually get good AI.
வாழ்த்துகள் Murai Labs team @murailabs @Mechramc
A Tamil-First Model, Trained from scratch, runs offline on a laptop, fully open. IMO small focused models like this are how Indian languages actually get good AI.
வாழ்த்துகள் Murai Labs team @murailabs @Mechramc
Agaram 2.5B is out. It's a Tamil-first language model we trained from scratch at @MuraiLabsHQ , our small non-profit lab. As far as we know, it's the first Tamil model trained from scratch that can call tools.
It chats in Tamil, Tanglish and English, translates, and can call the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop.
The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records.
Thank you to everyone who spent time with that early version.
The video walks through the model. You can try it here:
hf.co/spaces/mechram…
நன்றி 🙏
0
12
139
5,255
Ramchand Kumaresan retweeted
வாழ்த்துக்கள் @Mechramc!
தமிழ் மொழியைப் போல அகரமும் செழித்தோங்கி, காலத்தைக் கடந்து தனிச்சிறப்புடன் நிலைத்து நிற்கட்டும்!
தமிழ் மொழியைப் போல அகரமும் செழித்தோங்கி, காலத்தைக் கடந்து தனிச்சிறப்புடன் நிலைத்து நிற்கட்டும்!
0
1
1
163
Ramchand Kumaresan retweeted
@CMOTamilnadu Pls make it note of it and collaborate with @MuraiLabsHQ as this Tamil first language model will help taken government digital services delivery more effective and Interactive @itntHub @Akshita_N @TNeGA_Official
Agaram 2.5B is out. It's a Tamil-first language model we trained from scratch at @MuraiLabsHQ , our small non-profit lab. As far as we know, it's the first Tamil model trained from scratch that can call tools.
It chats in Tamil, Tanglish and English, translates, and can call the base model, chat model and GGUF files under Apache-2.0. You can download it and run it offline on your own laptop.
The first chat model had a four-day public preview. Seeing where it fell short with real users helped us decide what to work on. We rebuilt the fine-tuning around what we saw, using 222,662 records.
Thank you to everyone who spent time with that early version.
The video walks through the model. You can try it here:
hf.co/spaces/mechram…
நன்றி 🙏
3
4
17
611
5/ Two volunteers, Madhusudhanan G and Kaushik Kezendirane, contributed to this work and are co-authors of the paper. We're grateful for the time they gave it.
The paper documents the data, training and evaluation. We wrote it as a playbook for people working on other under-represented languages, including what worked and where we fell short.
The models and code are open for anyone to use, study and build on. Thank you for taking a look.
Weights: huggingface.co/murai-labs
GGUF (llama.cpp): huggingface.co/murai-labs/aga…
Code: github.com/Murai-Labs/Aga…
Paper: arxiv.org/abs/2610.18420
SFT Dataset: will be released soon
The paper documents the data, training and evaluation. We wrote it as a playbook for people working on other under-represented languages, including what worked and where we fell short.
The models and code are open for anyone to use, study and build on. Thank you for taking a look.
Weights: huggingface.co/murai-labs
GGUF (llama.cpp): huggingface.co/murai-labs/aga…
Code: github.com/Murai-Labs/Aga…
Paper: arxiv.org/abs/2610.18420
SFT Dataset: will be released soon
2
4
24
1,869
Same fix for the links above, written out in full:
Weights: huggingface.co/murai-labs
GGUF (llama.cpp): huggingface.co/murai-labs/aga…
Code: github.com/Murai-Labs/Aga…
Paper: arxiv.org/abs/2610.18420
Weights: huggingface.co/murai-labs
GGUF (llama.cpp): huggingface.co/murai-labs/aga…
Code: github.com/Murai-Labs/Aga…
Paper: arxiv.org/abs/2610.18420
6
1
19
1,622
Correct demo link (the one above opens as http:// in some browsers and can show a security warning):
huggingface.co/spaces/mechram…
huggingface.co/spaces/mechram…
2
0
15
1,305