Chef_AI (@itschef_ai) header
Chef_AI (@itschef_ai) profile picture

Mktg @SarvamAI | Always be cookin

Inside An LLM
Joined February 2015
Photo by Chef_AI (@itschef_ai)Photo by Chef_AI (@itschef_ai)Photo by Chef_AI (@itschef_ai)Photo by Chef_AI (@itschef_ai)Photo by Chef_AI (@itschef_ai)Photo by Chef_AI (@itschef_ai)Photo by Chef_AI (@itschef_ai)Photo by Chef_AI (@itschef_ai)Photo by Chef_AI (@itschef_ai)Photo by Chef_AI (@itschef_ai)
Slightly delayed, but I joined @SarvamAI's marketing team , and Token Factory is the first project I’ve worked on here. I benchmarked different providers on TPS and found that Sarvam’s token factory does really well. Things have moved really fast, and what we’ve achieved in such a short time seems unbelievable.🧵
Today, we are launching Sarvam Token Factory for everyone. We host our own models, and the full range of best-in-class open-weight models, in India. We match leading inference providers on performance and reliability. For enterprises, this means more models to choose from, and an AI stack they can own without lock-ins.
6
5
54
2,433
made this reel for @heykiviai ‘s Instagram and I love it😜
10
1
1
44
2,211
Picking a deepseek model provider these days is like finding a dealer. You never know what theyre cutting it with behind the scenes. I’ve seen the “same” Deepseek checkpoint come back with broken thinking traces, sometimes 3x tokens on same prompts, and other times giant latency increases just from switching between legit providers
9
2
12
1,411
Chef_AI retweeted
Nine open-weight models, one house, cameras everywhere. Full use of free will. modelbigboss.online/
3
2
1
24
1,544
the abuse of the word pareto by all AI companies shud be studied.
1
89
Chef_AI retweeted
what sarvam’s announcement taught us 1. Dudes have no clue about capex for a frontier/specialized lab 2. Dudes have no clue about inference-first as a business model for a lab
How is this not a fraud? India's leading AI lab worth $1.4 billion got funded and GPU to build a sovereign model in India by the Government. They have now stopped innovating. ..and using the same GPU to sell Chinese model openly. This is absolutely rubbish and mockery.
33
36
2
516
20,875
Chef_AI retweeted
Your feedback has shaped @SarvamAI's Model APIs for the last year, and now it will shape the Sarvam Token Factory over the next few days🚀 Every LLM on the Token Factory works with @livekit voice agents. I built one in 5 minutes: • Saaras v4 hears, Gemma 4 31B thinks, Bulbul v4 speaks • One OpenAI-compatible API, billed in ₹ • Served from India with zero data retention on request Looking forward to seeing what folks build and please let us know which integrations you would like us to prioritise next! Docs: docs.sarvam.ai Try it: indus.sarvam.ai
Today, we are launching Sarvam Token Factory for everyone. We host our own models, and the full range of best-in-class open-weight models, in India. We match leading inference providers on performance and reliability. For enterprises, this means more models to choose from, and an AI stack they can own without lock-ins.
15
12
2
81
2,827
There shud be an agent to track all e-lafda’s on tech twitter. Getting hard for someone with a job to stay updated with the drama.
3
7
1,239
Token Factory is finally out ‼️ I have been waiting for this launch for a long while now that it's finally out. I really love GLM and DeepSeek. They are so cheap, efficient, and fast. It's just too good for everything. While it was internally launched months back, I have used it to automate a lot of my workflows, from using Hermes as my daily driver, being driven by these models to a lot of small harness and workflows I have built internally which help me ship faster Go try out now!!
Today, we are launching Sarvam Token Factory for everyone. We host our own models, and the full range of best-in-class open-weight models, in India. We match leading inference providers on performance and reliability. For enterprises, this means more models to choose from, and an AI stack they can own without lock-ins.
17
20
3
208
8,323
5/ Moving state between GPUs. GLM 5.3 Flash’s draft cache reached the decode GPU only partly, cutting accepted tokens per step from ~5 to 1. Matched layouts, back to 4.7.
1
1
120
6/ I’ve had a lot of fun working with the team getting this out. Will make the repo for TPS arena public if enough people want to reproduce this (i've used the standard tier for all other providers and can share more detailed results as well). Feel free to DM me if you need help in comparing workloads, getting started, or anything else around the Token Factory. Get Started: docs.sarvam.ai/api/token-fac… Technical Breakdown: sarvam.ai/blogs/sarvam-token…
1
112
Chef_AI retweeted
Today, we are launching Sarvam Token Factory for everyone. We host our own models, and the full range of best-in-class open-weight models, in India. We match leading inference providers on performance and reliability. For enterprises, this means more models to choose from, and an AI stack they can own without lock-ins.
234
354
272
2,995
916,907
At this point I should just spin up a list of the people who prove theyve upgraded their inference engine against @vllm_project . From what I can see they’re all mostly a single prompt pony. I don’t see the same TPS throughout. I have no idea why the OSS community is not calling this out alr.
1
139
Is Inkling2 the response @thinkymachines has from the USA against Chinese open models? Maybe @thinkymachines should up it’s PR/influencer game and get folks to push more benchmarks. Not as much as @Kimi_Moonshot though😂
99
Replying to @alapanchaudhuri
There was an ICLR2026 blog post talking about this sort of thing: iclr-blogposts.github.io/202…
1
1
7
607
Spent the weekend reading this. I knew this but in all honesty hadn’t ever practically tried it out. It’s fun.
I am not sure how openly known this is but: you can read much of a model's training history straight from its weights. Like what? 1. which tokens were in the data 2. the LR schedule 3. the number of steps 4. and if you have the data: the step most documents were used in, the shuffle seed, the context window, the number of passes Some possibly less known facts that are verifiable this way (?): 1. In OLMo 2's first instruct release, 20,341 embedding rows that should have trained never moved, among them the token that closes every chat-turn marker. @allen_ai later found the tokenizer mismatch and retrained. The weights had it all along (👀@natolambert). 2. Pythia-1B was pre-trained at a peak learning rate of 2.5e-4, not the 3.0e-4 on its model card. Its training config agrees with the weights. 3. Pythia's learning rate sat flat for its last 28,000 steps: a cosine aimed at zero and clipped at a tenth of the peak. That is not the scheduler in today's gpt-neox. 4. Qwen2.5's instruct tuning touched exactly 4 of its 22 control tokens, the same four at 7B, 14B, 32B and 72B. The end-of-text, vision and fill-in-the-middle tokens show no trace of it. Pythia-1B model card (says 3.0e-4): huggingface.co/EleutherAI/py… Its training config (says 0.00025): github.com/EleutherAI/pythia… Today's gpt-neox scheduler: github.com/EleutherAI/gpt-ne… OLMo 2 SFT card, with AI2's note on the retrain: huggingface.co/allenai/OLMo-… OLMo 2 training data: huggingface.co/datasets/alle… Qwen2.5 base and Instruct: huggingface.co/Qwen/Qwen2.5-… huggingface.co/Qwen/Qwen2.5-… [1/2]
84
Gold fr
Crazy, I've just learned that Ross Anderson's Security Engineering (3rd edition) can be read for free (all chapters + full PDF: cl.cam.ac.uk/archive/rja14/b…) It's one of the most respected books on how to design secure systems. Definitely worth adding to your shelf, anon.
75