Oleksii Kuchaiev
Campbell, California, United States
9K followers
500+ connections
About
Leading AI model alignment and customization applied research at NVIDIA.
Activity
9K followers
-
Oleksii Kuchaiev reposted thisOleksii Kuchaiev reposted thisNeMo RL 0.6.0 is out with some killer features. Please check it out. - SGlang backend, Muon Optimizer, Speculative Decoding, Yarn long-context training, Chunked Cross Entropy Loss, top-p/top-k training, LoRA (GRPO/SFT/DPO), Fault tolerance launcher - Newly published SWE RL release benchmark demonstrating a long-context, multi-step RL rollout. https://lnkd.in/g5NYGpcu https://lnkd.in/gvXVnBJ7
-
Oleksii Kuchaiev reposted thisSpeculative decoding is a widely used technique for speeding up LLM inference at scale. This work explores how the same principles can be applied to accelerate RL post-training. Check this out! Paper: https://lnkd.in/gnYQpKNi Blog: https://lnkd.in/g5ES_reH Code: https://lnkd.in/gz8BR-YVOleksii Kuchaiev reposted thisRL post-training is hitting a rollout bottleneck. This new paper from #NVIDIAResearch shows how speculative decoding in NeMo-RL + vLLM can accelerate rollouts losslessly, with 1.8x higher throughput at 8B and projected 2.5x end-to-end speedup at 235B. Read the full paper: https://nvda.ws/4t8gPcw
-
Oleksii Kuchaiev reposted thisOleksii Kuchaiev reposted thisNVIDIA Nemotron 3 Nano Omni is OUT. WE GAVE NEMOTRON EYES AND EARS. Best part, for me anyway: it is not another stitched-together stack where one model looks, one model listens, one model reads, and then everybody drops context on the floor between turns. It is one open model handling the whole messy multimodal situation. Audio. Video. Images. Text. Native audio support is a huge add. 256K context is great. The benchmark story looks strong. The throughput story looks strong too. And we put out weights, recipes, plus data. Hell, you could build your own Omni if you wanted! This model can be extremely useful for real agents as a subagent that can reasoning through complex documents and videos. Read the paper, read the tech blog, read the recipes, look at the data - we're releasing everything as per usual!
-
Oleksii Kuchaiev shared thisNemotron 3 model family keeps growing with new members!Oleksii Kuchaiev shared thisIntroducing NVIDIA Nemotron 3 Nano Omni, an open multimodal model that unifies video, audio, image and text reasoning within a single model. Built for agentic AI workloads, it helps developers build up to 9x more efficient AI agents with leading multimodal accuracy.NVIDIA Launches Nemotron 3 Nano Omni Model, Unifying Vision, Audio and Language for up to 9x More Efficient AI AgentsNVIDIA Launches Nemotron 3 Nano Omni Model, Unifying Vision, Audio and Language for up to 9x More Efficient AI Agents
-
Oleksii Kuchaiev reposted thisOleksii Kuchaiev reposted thisThank you to everyone in the community who is testing and using Nemotron models. It's great to see Nemotron-Cascade-2, Nemotron-3-Super and Nemotron-3-Nano trending on HF. The Nemotron team is working hard to incorporate all your feedback into Nemotron 4. And yes, Nemotron 3 Ultra is still on track for release. https://lnkd.in/gQM2a_gU
-
Oleksii Kuchaiev reposted thisOleksii Kuchaiev reposted thisReasoning benchmarks are vital for measuring progress on structured tasks and when we share methods openly, the entire community moves faster. To put this into practice, we’re excited to announce the NVIDIA Nemotron Model Reasoning Challenge hosted by NVIDIA AI and powered by Google Cloud Partners. Participants will start with a Nemotron-3 Nano baseline and a novel reasoning benchmark from NVIDIA Research. The goal is to develop techniques that push the boundaries of reasoning accuracy using open models. All compute runs on Google Cloud G4 VMs featuring NVIDIA RTX PRO 6000 Blackwell GPUs, giving participants high-performance infrastructure for fast iterations and evaluation. 💰 $106,388 Prize Pool ⏰ Entry Deadline: June 8, 2026 Ready to build? Learn more here: https://lnkd.in/dvbTJZzb
-
Oleksii Kuchaiev shared thisIf you are attending GTC 2026, stop by tomorrow, Tuesday, March 17 at 3:00 p.m. - 3:40 p.m. PDT to learn about how we post-train Nemotrons. Looking forward to seeing everyone! Session number: S81558
-
Oleksii Kuchaiev shared thisAnd Nemotron 3 Super is top open model for OpenClaw, according to https://pinchbench.com/
-
Oleksii Kuchaiev shared thisNemotron 3 Super is here — 120B total / 12B active, Hybrid SSM Latent MoE, designed for Blackwell. Truly open: permissive license, open data, open training infra.
-
Oleksii Kuchaiev liked thisOleksii Kuchaiev liked thisWe're launching the agentic robotics app store today. Let's democratize AI robotics for all! 300+ apps shipped. 10,000 robots in the wild. It used to take weeks from a robotics engineer to build apps, now everyone can do it in hours with ML intern or your favorite neighborhood agent! My favorite reachy mini app was built by Joel, a 78yo marketing exec who'd never coded in his life. Personally, I built an office receptionist in two hours last week. More info to start building here: https://lnkd.in/d4vRPbvt
-
Oleksii Kuchaiev liked thisOleksii Kuchaiev liked thisJust got back from an incredible week at #ICLR in Brazil! 🇧🇷 I had the opportunity to present some of NVIDIA’s Research (RLBFF & ProfBench, links below), and connect with collaborators & old colleagues. It was rewarding to hear first-hand accounts of how NVIDIA’s open-source datasets and models are being leveraged across the ecosystem.
-
Oleksii Kuchaiev liked thisPerplexity runs on NVIDIA. Nice breakdown from the team on how they’re using the CUTLASS Python stack to optimize their models for inference 👇Oleksii Kuchaiev liked thisWe’ve developed our own inference engine Runtime-Optimized Serving Engine (ROSE) to serve models ranging from embeddings to trillion-parameter LLMs. With CuTeDSL integrated into our inference engine, Perplexity can build the specialized GPU kernels faster to bring models up to peak performance on NVIDIA Hopper and Blackwell GPUs. Read the full post on our research blog. https://lnkd.in/gTeJxQEZ
-
Oleksii Kuchaiev liked thisOleksii Kuchaiev liked thisI'm excited to share that I’m starting my internship with the NVIDIA NeMo team as a Deep Learning Algorithm Engineering Intern! 🚀 I’ll be working alongside Bernard Nguyen and Yu Yao in Santa Clara, diving into cutting-edge AI infrastructure technologies. Excited to learn, build, and contribute to impactful work in the California Bay Area this summer. Looking forward to the journey ahead!
-
Oleksii Kuchaiev liked thisDeeply honored to receive the 2025 IEEE SPS Claude Shannon-Harry Nyquist Technical Achievement Award at ICASSP 2026, Barcelona. This is a significant recognition of the work I have conducted with my mentors, colleagues, and other collaborators over the decades.
-
Oleksii Kuchaiev liked thisOleksii Kuchaiev liked thisAttended an insightful online session on NVIDIA Nemotron 3 Nano Omni 🚀 — packed with practical learnings and future-focused ideas. Key takeaways ✨ • Multimodal capabilities unlocking smarter agent reasoning • Using sub-agents effectively in multi-agent systems • Choosing the right model for the right task • Open innovation through the Nemotron Coalition Big shoutout to the amazing speakers — Corey Noles, Chris Alexiuk & wendell wilson 🙌 Excited for what’s ahead in AI agents! 🤖✨ #AI #NVIDIA #Nemotron #AIAgents #Tech
-
Oleksii Kuchaiev liked thisOleksii Kuchaiev liked thisI've finally made the jump to foundation model development. The AV industry has been my home for my entire career so far, but it's time to leave that comfort zone behind. I've started a new chapter at NVIDIA as a Staff Research Engineer on the Nemotron Post-Training team, working with an incredible group of researchers and engineers behind our Nemotron models. After several years working on metrics and data-driven evaluation for autonomous vehicles, I'm really looking forward to moving deeper into the LLM stack — focusing on the RL and post-training recipes that turn raw pre-trained models into useful, aligned, and reasoning-capable systems. I hope to be contributing to NVIDIA's open-source post-training infrastructure (NeMo-RL, NeMo-Gym, NeMo Framework) and tackling the end-to-end challenges of large-scale training and inference.
Experience
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
-
Education
-
-
-
-
-
-
Licenses & Certifications
-
Executive Data Science Specialization
Coursera Course Certificates
-
Fundamentals of Management
Coursera Verified Certificates
-
-
Functional Programming Principles in Scala
Coursera Verified Certificates
Languages
-
English
Full professional proficiency
-
Russian
Native or bilingual proficiency
Recommendations received
4 people have recommended Oleksii
Join now to viewOther similar profiles
Explore more posts
-
Pierre de Lacaze
Reinforcement Pre-Training (Microsoft, June 2025) Paper: https://lnkd.in/eY4PCpUi Abstract: "In this work, we introduce Reinforcement Pre-Training (RPT) as a new scaling paradigm for large language models and reinforcement learning (RL). Specifically, we reframe next-token prediction as a reasoning task trained using RL, where it receives verifiable rewards for correctly predicting the next token for a given context. RPT offers a scalable method to leverage vast amounts of text data for general-purpose RL, rather than relying on domain-specific annotated answers. By incentivizing the capability of next-token reasoning, RPT significantly improves the language modeling accuracy of predicting the next tokens. Moreover, RPT provides a strong pre-trained foundation for further reinforcement fine-tuning. The scaling curves show that increased training compute consistently improves the next-token prediction accuracy. The results position RPT as an effective and promising scaling paradigm to advance language model pre-training."
1 Comment -
Jeremy Fowers
Are mixture of expert (MOE) models the future for professionals using local LLMs? Buying enough VRAM for 30-100B MOEs on a Strix Halo mini PC is quite affordable, and the inference speed for the 30B MOE is just awesome. Filmed on Ryzen AI MAX 395+ with 128GB RAM, Vulkan backend, using Lemonade: https://lnkd.in/eDZW7ZZn Great work by Daniel Holanda Noronha sprucing up the interface! Ramakrishnan Sivakumar Victoria Godsoe Tomasz Iniewicz Kalin Ovtcharov
2 Comments -
Osama Jaber
Forge delivers 2.4–5.2× faster inference than torch.compile on NVIDIA B200 We benchmarked Forge against torch.compile(mode='max-autotune-no-cudagraphs') the strongest PyTorch baseline available across 7 production models Now publicly available with a free trial: https://lnkd.in/dnw_XmJH
2 Comments -
Devesh Raj
Architectural breakthroughs in Gen AI continue! Moonshot AI’s Kimi K2 has achieved top leaderboard positions, surpassing GPT-4, thanks to key innovations—especially the introduction of the MuonClip optimizer. This new approach stabilizes training for massive Mixture-of-Experts (MoE) models, significantly boosting performance and efficiency at unprecedented scale. https://lnkd.in/gvYrGzTK
-
Ravi Annaswamy
https://lnkd.in/gHztfcgY The best way to interpret models is to train interpretable models! Large neural networks learn mapping functions that translate vectors to vectors (vector sequences to vector sequences in the case of LLMs and Seq2Seq machines). If that statement sounds dense, let me explain. An image classifier takes a 256x256 image of a dog and spits out 'Dalmation' as its class label. Input was a flat 65536 (256 times 256 pixels of color values) vector. Output was a vector of 1000-possible objects with dalmation with value close to 1.0, horse with value close to say 0.5 and ship with value close to 0.1 and ocean close to 0! Neural networks learn such complex mappings by learning to compute features on features - mathematically they learn a series of rotations of a vector with trimming/clipping. They learn correct mappings but one cannot follow the features being computed at each step. Two models can learn mappings that are FUNCTIONALLY equivalent, but not comparable step by step! How do we then interpret HOW the models reached their decision? == Work from @OpenAI shows that forcing models to have most weights as zero, forces them to build sparse, modular circuits. An example shows how a network keeps track of whether 'Hello will be followed by ' or ". Some of the most interesting work along these lines - were from @tegmark and @ZimingLiu11 - training sparse modular networks. BIMT Biologically inspired Modular Training - if my memory serves right. In contrast, mechanistic interpretibility work can probe and interfere, and instrument ANY trained network using sparse auto encoder maps to see what is happening. Anthropic @ChrisOlah and DeepMind @NeelNanda5 have some outstanding insights along those lines. That shows that EVEN when not incentivized to be sparse and modular, the learning, after grokking happens, leads to circuits that can be distilled to sparse modular circuits. The interpretable-by-design is fascinating bottom up way and the MI on complex REAL large models is fascinating in insights. I think the both might converge soon feeding insights to each other.
2 Comments
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content