Stas Kirillov

Generating tokens in Nebius Token Factory, ex catboost, ex Yandex

[Stas Kirillov @kizillof]

@ClementDelangue Looks super! Are you planning to offer a fully or partially disassembled version, so assembling it with the kids would give additional fun?

[Stas Kirillov @kizillof]

@yeshuaige9988 @nebiustf @nvidia @Alibaba_Qwen On their way! Will be there soon.

[Stas Kirillov @kizillof]

@MaximBlue23 Yes, fixed this mostly both in instruct & Thinking, so give it a try! Also Kimi-K2-Instruct-0905 finally also will be available soon.

[Stas Kirillov @kizillof]

Mine first release as a part of Nebius Token Factory team, please send feedback if any :)

[Nebius Token Factory @nebiustf]

Kimi K2 Thinking from @Kimi_Moonshot is live on Nebius Token Factory!

Currently the strongest open-weight model on the market and competitive with proprietary stacks, with full 262K context.

Key highlights:

• SOTA reasoning among all open 70B models (beats or matches GPT-5.1/DeepSeek/Qwen in many tasks)
• Reliable long-context performance, 262K that actually reasons
• 3–5× faster inference than comparable frontier models
• Top-tier math & code results (AIME, MATH-500, LiveCodeBench)
• Near-zero refusal rate + strong multilingual output

Run it today → https://t.co/1pHiVpV6oW

[Stas Kirillov @kizillof]

@SpV_coin @nebiustf @PrimeIntellect Today 😊

[Nebius Token Factory @nebiustf]

⚡️Today marks a big milestone for Nebius.

We’re launching Nebius Token Factory, the evolution of Nebius AI Studio, built to make open-source AI production-grade.

Token Factory transforms raw open models into governed, scalable systems with dedicated inference, sub-second latency, 99.9% uptime and zero-retention compliance.

It’s where inference, post-training and governance converge, turning raw compute into reliable intelligence.

Run AI inference at scale: https://t.co/rRIK3rpaLI

[Stas Kirillov @kizillof]

@playratewinrate Hi! Did you measure logs + models or just resulting model files? In CatBoost, we log lot's of data in different formats by default - human readable train log, train/validation metrics and stats in json format for each iteration and so on.

[Stas Kirillov @kizillof]

@takes_mediocre I'm so sorry! Released python wheel will be enough?

[Stas Kirillov @kizillof]

@tmarzagao Next friday we'll finally publish 1.0.5 with m1 & x86_64 universal binaries for Mac os. Sorry for such a long wait 😸

[Anton Tsitsulin @graph_]

Only in Moscow there are billboard ads for Python packages.

Bonus: Moscow State’s main building in the background.

[Stas Kirillov @kizillof]

@Yasushi_ML Hello! We are currently refactoring documentation pages, can you help us to find, how to make it better? You can either write here, or leave feedback on documentation pages with "thumbs down" button in the upper right corner of page and leave your comments.

[Stas Kirillov @kizillof]

@woman_ess Привет! А расскажи, это CPU или GPU обучение? Много данных?

[Knut Jägersberg @JagersbergKnut]

For heterogeneous data, @CatBoostML is the best high performing baseline model to grab!

https://t.co/rclg0ZaB3w

[Stas Kirillov @kizillof]

@StefvandenElzen @DariusAtmar @DataSciNews Hello! CatBoost supports plotting single tree with graphviz, so you can look at plot_tree() https://t.co/tKkyCZbJr1 function realization as reference. Important note: CatBoost has two different tree types - symmetric(aka oblivious) and nonsymmetric. I can help you with that 😺

[Stas Kirillov @kizillof]

@JagersbergKnut By no tuning we mean reasonable quality out of the box (that's good for quick start), but tuning even simple options like learning rate & number of trees can significantly change final quality.

[Nikita Pavlichenko @nv_pavlichenko]

I'm excited to share my first paper accepted to NeurIPS @NeurIPSConf. This work was a culmination of the long path we did at @TolokaAI of improving crowdsourced text aggregation.

Reach us at NeurIPS or check out our paper at arXiv: https://t.co/6gF6uYvdkz

1/N

[Stas Kirillov @kizillof]

@planarrowspace Yes, parameter named grow_policy described here : https://t.co/4WJ53zZUjW

[Stas Kirillov @kizillof]

@Oktahedro @CatBoostML Great! You can start here, in main export place: https://t.co/kxumDbi0De and then go to helper functions implementation defined in https://t.co/B9e5RPDNEj and in corresponding header.
Read about building python version here: https://t.co/QijYfDGCHH 🐈

[Stas Kirillov @kizillof]

@abap34 Yes! Since 1.0.0 version CatBoost supports multilabel classification. We are waiting for documentation update, corresponding loss names are MultiLogloss and MultiCrossEntropy