Soham Chatterjee

Co-Founder and CTO @scaledown_team
Building Task Specific Small Language Models

[ScaleDown @scaledown_team]

NYC builders - we're co-hosting a Voice AI Agents Hackathon with @AgoraIO on June 26–27 in Midtown Manhattan

100+ engineers in one room

Build a voice agent, ship it, and prove it's the best

$4.5K prize pool 🏆

🔗 Register 👉 https://t.co/HmFLmBkK5M

[Neal Patel @neal_k_patel]

OpenAI ☑️ Anthropic ☑️ Google 👀

@scaledown_team's task-specific SLMs beat Gemini 3.5 Flash (Google's most intelligent model built for speed):

Accuracy: 11.48% higher
Speed: 10.5x faster
Cost: 36x cheaper

Full benchmarks: https://t.co/A2ikV8s32k

[Soham Chatterjee @csoham_]

@davids_works @neal_k_patel Hey @davids_works The context lenght is unlimited for the compression model and 1M tokens for summarize, classify and extract endpoints.
Yes, we can extract sentiment! That's one of the most common usecases people use us for! We also return the reasoning with the classification!

[Soham Chatterjee @csoham_]

@asatoucan @neal_k_patel Hey @asatoucan It's actually not an orchestrator or router. We have our own models for these tasks. We also have an MCP server that you can add to your coding agents!

[Soham Chatterjee @csoham_]

@itsArmanj @neal_k_patel Hey @itsArmanj That's something we've heard a lot from our customers when we started. Our benchmarks are on OOD data with no finetuning. We wouldn't launch if we didn't solve this problem. Finetuning is only needed for niche domains like legal or pharma.

[Neal Patel @neal_k_patel]

Introducing ScaleDown

15x cheaper. 63x faster. 5.1% more accurate than GPT-5.4 Mini.

Task-specific SLMs for the 70-80% of AI workloads that don't need a frontier model.

From NeurIPS '25 to https://t.co/imZLOYCcTA

[Aaditya Ura @aadityaura]

Introducing OpenBioLLM-Llama3-70B & 8B: The most capable openly available Medical-domain LLMs to date! 🩺💊🧬

Outperforms industry giants like GPT-4, Gemini, Meditron-70B, Med-PaLM-1, and Med-PaLM-2 in the biomedical domain. 🏥📈 🌟

OpenBioLLM-70B delivers SOTA performance, setting a new state-of-the-art for models of its size.
OpenBioLLM-8B model and even surpasses GPT-3.5, Gemini, and Meditron-70B! 🚀

Today's release is just the beginning! In the coming months, we'll be introducing:

- Expanded medical domain coverage 🧠
- Longer context windows 📜🔍
- Better benchmarks 📈🏆
- Multimodal capabilities🖥️🩺📊🔬

Medical-LLM Leaderboard: https://t.co/rSM98TLidd

#gpt #gpt4 #gemini #medical #llm #chatgpt #opensource #llama3 #meta

[Layoff Support @Laidoffsupport]

Happy Monday! 🌟 Remember, transitions are periods of growth. Take time to heal, reflect, and reach out. If you're navigating a career change, you're not alone. Layoff Support is here to guide you. #MindfulMonday #CareerGrowth #LayoffSupport
https://t.co/AiBLn5eFYw

[Layoff Support @Laidoffsupport]

Thinking about your next career move? Consider VC-backed startups! High growth, impactful work, and the chance to be part of something big. This week, we're sharing why and how to dive in. 🌈💼 Explore VC job boards on Layoff Support! #CareerLeap #VCJobs #StartUpLife

[Soham Chatterjee @csoham_]

"Hey Siri, but make it a scientist! 🧬👩‍🔬 Meet #GeneGPT, your go-to AI for midnight research marathons. #ScienceMeetsSiri"
https://t.co/6Z2H4h6d4i

[WiML @WiMLworkshop]

🎉 What an incredible journey! Big thanks to everyone who joined the 18th WiML Workshop at NeurIPS. Special shoutout to our speakers, panelists, mentors, and affinity groups for enriching discussions and collaborations. And our sponsors, you rock! #NeurIPS #WiMLWorkshop

[Soham Chatterjee @csoham_]

The latest Bard release powererd by PaLM 2 is better, but still leaves a lot to be desired when compared to GPT-4

#BardAI #BardByGoogle #GPT4 #ChatGPT #LLMs

[Archana Vaidheeswaran @varchanaiyer_]

🧵I've been hearing all about Google Bard for the past few days, so I decided to put it to the test. I compared Bard with ChatGPT using "everyday prompts", and here's what I discovered. 👇

[Soham Chatterjee @csoham_]

@xuanalogue Our finetuned prompts for 002 dropped in performance for 003. Unfortunately, since 003 came out, the latency for 002 increased by quite a lot (>1min in some cases). So if you are someone building a product, you have no choice but to use 003

https://t.co/jKKtxcMF6e

[(((ل()(ل() 'yoav))))👾 @yoavgo]

i asked autogpt to [do something simple] and it did an amazing job! (which i didnt verify). i am sure that if i asked it to [do a much more complex version of it] it would do a great job and be amazing!!

[Archana Vaidheeswaran @varchanaiyer_]

🔥 Is building NLP products too easy now? 🚀

Our latest newsletter has a fresh take on this topic that you won't want to miss. Discover the challenges and ideal applications of Large Language Models (LLMs) in our Chrome extension blog. Plus, get an inside look at our backend architecture and ways to contribute to our open-source project.
https://t.co/Gz2o8Z1c8Y

[Soham Chatterjee @csoham_]

@HoddieOt Thanks for sharing! Glad you found the article insightful!

[Artem Andreenko @miolini]

I've sucefully runned LLaMA 7B model on my 4GB RAM Raspberry Pi 4. It's super slow about 10sec/token. But it looks we can run powerful cognitive pipelines on a cheap hardware.

[Stacey Higginbotham @gigastacey]

Stacey on IoT: Why is the browser a second-class citizen in smart homes?: https://t.co/cDNCSBnGu8

[Eli Hughes @emh203]

Hubris in embedded systems is a liability.

Always start here. No exceptions.

[Alasdair Allan @aallan]

If you've been following our progress on the #Picamera2 Python camera library you'll be happy to know that last week's #RaspberryPi OS update included a preinstalled beta release library. Manual is out of draft and can be found at https://t.co/LzDqwjd9N5. https://t.co/Ys0aE1wiUd