Daanish Khazi

We’re announcing Kos-1 Lite, a medical model that achieves SOTA on HealthBench Hard at 46.6%. As a medium sized language model (~100B), it achieves these results at a fraction of the serving cost of frontier trillion-parameter models.

@madiator pip install rubric github.com/The-LLM-Data-C…

What's the library people use for defining/loading/processing rubrics?

@AyushKarupakula such a great read!

Excited to finally share Ebla-1 and the C⁴ benchmark. Really enjoyed working with HUD on the evals behind it.

labs will publish details on arch, optim, objectives, scaling, kernels, literally everything except data and academia will be astounded for the hundredth time, wondering to itself where the secret sauce is

@Szhu4433 Dmed you!

@bertgodel This is so cool! Password please

We’re announcing Kos-1 Lite, a medical model that achieves SOTA on HealthBench Hard at 46.6%. As a medium sized language model (~100B), it achieves these results at a fraction of the serving cost of frontier trillion-parameter models.

@iohcsnim @GoogleDeepMind Congrats Min!! 🌲🏠

One year at @GoogleDeepMind, and I'm so excited to share something I've worked on this past year! Introducing Gemini Embedding 2: our first natively multimodal embedding model that maps text🔠, images🖼, video🎞, audio🔊, and documents📄 into a single embedding space.

@arthursolwayne @DealGlassInc Congrats Arthur!

i quit my quant job at SIG to build @DealGlassInc we just topped SpreadsheetBench Verified, mogging Opus 4.6, GPT-5.4, and Shortcut. generic spreadsheet agents optimize for speed at the expense of quality. we built Tetra to underwrite billions.

thanks vamsi! it seems hard to have a policy that always does what it's told (a la IFBench) and also pushes back and resists sycophancy when necessary see gpt 5.4 system card today - they note degradation on HB Hard for specifically this: "GPT-5.4 seeks much less context than GPT-5.2.. its main weaknesses are poorer context-seeking when information may be missing" getting the model to eagerly complete tasks (and score high on SWE Bench) is likely at odds with getting it to clarify, push back or re-orient the user when necessary

@bertgodel Couldn't an LLM be concise, compassionate and accurate - but also do good Instruction Following at the same time? Not sure why they should be different

@lightc0n3 yeah definitely

@bertgodel Can we put Kos-1 Lite on EndpointArena.com to see how it does at predicting FDA outcomes? @bertgodel

@ketcholito Thank so much Annalise!!

This team is so good. There are <20 companies in the US who can effectively post train a 100b MoE model, and far fewer at the SOTA level. Can't wait to see what is next

@thegavinbains LMAO

data gf 🤝 compute bf

@MiasBear dmed you!

@bertgodel @bertgodel I can't direct message to you, please send me a password for kos-1 lite demo, Thank you!

@YatinBadal dmed you!

@bertgodel Very cool! I’d love to try it too - what’s the password?

@knowledge_embed dmed you!

@bertgodel Can I get the password? I want to try it

@salathe dmed you!

@bertgodel Password please

@MichaelN_RS dmed you!

@bertgodel Password please!

@meehowee dmed you!

@bertgodel Awesome! I'm a cardiologist/pharmacist working with AI - I would love to check it out, if possible. Please let me know 🙂

@mercier_kent dmed you!

@bertgodel Please reply with a password to try out Kos-1 Lite! I found out about this from the March 04, 2026 newsletter by Dr. Alex Wissner-Gross. Thanks, can't wait to try it.

@JohnsonThomasMD dmed you!

@bertgodel How do I get access. Can’t DM you 😞

Try Kos-1 Lite here: kos.llmdata.com Read the full blog: llmdata.com/blog/kos-1

We're already seeing strong reactions from physicians in early testing, and we're excited to get this in front of more clinicians and users. We are continuing to improve and train the model, and we’ll release upgraded checkpoints as training continues.