NVIDIA Nemotron-Cascade 2: MoE with 30B parameters and gold medals in math competitions.

Nemotron-Cascade 2 - world's second open language model that won gold medals in three competitions in 2025: IMO, IOI, and the ICPC finals.

➡️ Why it matters?

Before it, only DeepSeek-V3.2-Speciale, a model with 671B parameters and 37B active ones, had achieved this. Nemotron-Cascade 2 has 20 times fewer parameters: 30B total, 3B active.

☞ At IMO 2025, the model solved 5 out of 6 tasks and scored 35 out of 42 points.
☞ At IOI - 439 out of 600, corresponding to gold.
☞ At ICPC World Finals, 10 out of 12 tasks were solved, ranking 4th in the gold zone. Eight ICPC tasks were solved in less than 100 attempts.

➜ Architecture

MoExperts based on Nemotron-3-Nano-30B-A3B-Base. The difference from the first version is an expanded Cascade RL, where reinforcement learning takes place sequentially across domains: first following instructions, then multi-domain RL, then distillation, RLHF, working with long context, code, and finally software engineering tasks.

➜ Multi-Domain On-Policy Distillation (MOPD)

At each stage of Cascade RL, the best intermediate checkpoint for a specific domain is selected and used as a teacher. Since all teachers come from a single SFT initialization, they share the tokenizer and vocabulary, which simplifies distillation.

MOPD works at the level of individual tokens, not sequences, and converges faster than classic GRPO: on AIME 2025, it reaches the teacher's level in 30 optimization steps.

➜ TESTS

On benchmarks, the model outperforms Qwen3.5-35B-A3B and the larger Nemotron-3-Super-120B-A12B in math, code, and following instructions.

☞ LiveCodeBench v6: 88.4 (Qwen3.5 - 74.6);
☞ ArenaHard v2: 83.5 vs. Qwen3.5's 65.4;
☞ IFBench: 82.9 vs. 70.2;
☞ On Codeforces, the model scored a rating of 2345 with TIR (on par with models with 300B+ parameters).

➜ WEAKNESSES

Tasks requiring deep knowledge (MMLU-Pro, GPQA-Diamond) and agent scenarios (BFCL v4, τ²-Bench, SWE Verified). Here, Qwen3.5 is still ahead.

NVIDIA promises to improve the science-intensive pretraining and agent RL in future versions of Nemotron-Cascade.

Licensing: NVIDIA Open Model License: Model, SFT dataset, RL data, Technical report

#AI #ML #LLM #Nemotron #Cascade2 #NVIDIA

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

BitVLA: 1-bit VLA Models for Robotics Manipulation

An interesting study on Vision-Language-Action (VLA) models has been released - this is AI that allows a robot to see an object, understand a task, and act (for example, grab an item).

Usually, this requires huge models and expensive GPUs.
But here,
➡️ How they did it differently?

The team developed a model where almost all calculations are reduced to 3 values:
-1, 0, and 1

Yes, literally.

Thanks to this:
- the model was compressed to ~1.4GB
- they removed complex mathematics
- they made it suitable for cheap hardware

👉 the accuracy remained at the level of large models
👉 the speed increased by 4 times

That is, the robot moves just as accurately, but faster and cheaper:

- expensive servers and GPUs are no longer needed
- robots can be run directly on local chips
- the path to mass robotization (at low cost) is opened up

Paper

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Jacobian is how multidimensional systems actually "move".

You're no longer working with a single variable —
you're working with transformations:

input vector → output vector

➡️ Jacobian shows how each input dimension affects each output.

What it is?
→ a matrix of partial derivatives
→ each row = one output function
→ each column = one input variable

J(i,j) = ∂fᵢ / ∂xⱼ

Why it's important?
→ it's a local linear approximation of a non-linear system
→ it shows how small changes propagate
→ it transforms complex systems into a computable form

In physics:
→ coordinate transformations
→ velocity mapping
→ variable substitution in integrals

In robotics:
→ maps joint velocities → end-effector speed
→ singularities occur when the Jacobian degenerates

In optimization / ML:
→ gradient propagation through layers
→ backprop is a chain of Jacobians

Interpretation:
→ det ≠ 0 → the transformation is locally invertible
→ det = 0 → there's a loss of information

Jacobian isn't just a theory.
It's an interface between geometry and computations.

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Chrome might become too heavy for the era of AI agents.

Lightpanda, a headless browser, written from scratch, which works 11X faster and uses 9X less memory

And most interesting thing is that it is not based on Chromium.

➡️ Why need of another?

Currently, almost all AI agents that automate work with websites use Chrome under the hood.

But there's a problem.

You're actually running a huge desktop application without an interface, sometimes hundreds of times on a server.

And all this for a task where there's no need to render a single pixel.

This is like renting a wagon to send a letter.

Lightpanda does everything differently.

It's written in Zig and was originally created for just one purpose -
the fastest headless browser for automation.

At the same time, it still supports:

• JavaScript
• Ajax
• Fetch / XHR
• SPA applications
• infinite scroll

But without hundreds of megabytes of browser "ballast", which you don't need.

It's also very easy to connect:

• works with Playwright, Puppeteer and chromedp via CDP
• one Docker command
• CDP server on port 9222
• can replace Chrome literally in 30 seconds

Where this is particularly useful:

• AI web agents
• collecting data for LLM training
• large-scale browser automation
• test pipelines

Project is still in beta, but has already gathered 11.8K stars on GitHub and it's open source under AGPL-3.0 license.

Seems that a new browser standard might emerge for AI agents.

GitHub

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Supermemory scored 99% on SOTA memory (LongMemEval_s) benchmark Using experimental ASMR (Agentic Search and Memory Retrieval) technique

☞ Instead of classic vector search and embeddings, a system of parallel agent-observers is used, which extract structured knowledge from the history of interactions in several directions at once

☞ Specialized search agents are used:
→ for exact facts
→ for related context
→ for restoring the temporal sequence

☞ And most interesting thing is that no vector database is required

The project will become open-source in just 19 days, let's keep an eye on it.

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

🎧 Fish Audio S2 Pro is a powerful open-source model for speech synthesis

This is a 4B + 400M parameter TTS model with an unusual Dual-AR architecture, trained on 10+ million hours of audio and supporting 80+ languages.

🟢 What makes it interesting?

🏗 Dual-AR architecture

The model is divided into two parts:

• 4B Slow AR is responsible for the semantics and structure of speech
• 400M Fast AR generates 9 residual codebooks for acoustics

This approach allows to maintain high sound quality without a heavy load on inference.

🎭 Free control of emotions and intonation

You can directly set the style of speech in the text:

[whisper]
[laughing]
[professional broadcast tone]

15,000+ tags are supported, which work at the level of individual words.

🌐 80+ languages

The main high-quality languages:

• English
• Chinese
• Japanese

⚡ Optimization for LLM infrastructure

The model natively works with the SGLang stack, so it supports:

• continuous batching
• paged KV cache
• RadixAttention prefix caching

📊 Performance

• RTF: 0.195 on Nvidia H200
• ~100 ms to the first audio
• more than 3000 acoustic tokens/sec

Developers released: Model weights, code on GitHub for fine-tuning, Engine for streaming inference.

#AI #TTS
••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Autonomous AI research with autoresearch

This repository proposes concept of autonomous AI learning, where agent modifies code and conducts experiments on its own.

Key points:

- The autonomous agent modifies train.py to optimize the model.

- learning process takes place in a fixed timeframe of 5 minutes, after which it evaluates the results and continues the iterations.

- Using a simple interface program.md, users can configure the agent to optimize models without directly interfering with the code.
- Support for only one NVIDIA GPU.

GitHub #python

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

DRAGOn methodology.

Russian researchers from SberAI, MWS AI, ITMO and HSE University, have presented the new approach to evaluating RAG systems, which underpin modern AI assistants.

➡️ Why Most AI Assistant Tests Don't Work in Reality?

The key idea is to move away from static tests to a dynamic environment with constantly updated data. The work was accepted at the EACL 2026 international conference.

Classic benchmarks quickly become outdated and poorly reflect real-world conditions. In business, AI works with live knowledge bases, where the relevance and coherence of facts are important, not just accuracy on a fixed dataset. DRAGOn proposes to test AI systems on fresh news, automatically collecting a "knowledge map" from them.

Instead of simple questions like "who/where/when", the system creates multi-level logical tasks. To answer, AI must match several facts from different news, not just copy a piece of text, and the neural network-judge checks the answers.

What this brings in practice:
- Tasks become multi-step, not trivial;
- The ability to link facts, not copy answers is tested;
- The evaluation takes into account completeness and factual accuracy, not just word matching.

Methodology can be deployed within a company and test AI on its own data before implementation. This allows comparing solutions in real scenarios and reducing the risk of errors, especially in tasks of analytics, support, and working with documents.

Article

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Practice for PyTorch Interviews

TorchCode offers a structured environment for training programming skills needed for ML interviews. Solve problems implementing operators and architectures, receiving instant feedback and hints.

➡️ Key features:

- 40 tasks frequently encountered in interviews
- Automatic verification of correctness and performance
- Instant feedback on each test
- Hints and reference solutions for study
- Ability to run in the browser without installation

GitHub #Python

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Google developed a KV-cache quantization algorithm without loss of accuracy.

Research division announced TurboQuant, a vector quantization algorithm that combines 2 other methods - QJL and PolarQuant, which solves the problem of increasing KV-cache when working with a long context.

TurboQuant will be presented at ICLR 2026, PolarQuant - at AISTATS 2026.

➡️ Why it matters?

The KV-cache stores intermediate representations of tokens so that the model does not recalculate them at each step of generation. With the growth of the context, it becomes a memory bottleneck.

Conventional vector quantization compresses this data, but it introduces overhead: for each block, it is necessary to store quantization constants in full accuracy, and this is plus 1–2 bits per element, which partially devalues the compression itself.

➜ TurboQuant is a two-stage pipeline.

First, PolarQuant: a random rotation aligns the geometry of vectors, after which they are translated from Cartesian coordinates to polar (radius and angle). The distribution of angles turns out to be predictable and concentrated, so normalization and storage of additional constants are no longer necessary.

At the second stage, QJL is connected, a method based on the Johnson-Lindenstrauss transformation, which encodes the residual error of the first stage with just one sign bit and, through a built-in estimation function, combines a high-precision query with low-precision compressed data, correctly calculating the attention score.

None of the methods requires training or retraining and works in the mode of "without preliminary analysis of the dataset".

The algorithms were tested on benchmarks for long context: LongBench, Needle In A Haystack, ZeroSCROLLS, RULER and L-Eval with the Gemma and Mistral models.

When quantizing the KV-cache to 3 bits, TurboQuant showed zero degradation in accuracy on all tasks: searching for a "needle in a haystack", QA, code generation, summarization.

The volume of the KV-cache at the same time decreased by 6 times. On H100, the four-bit TurboQuant accelerated the calculation of attention logits up to 8 times compared to 32-bit keys.

Area of application is not limited to KV-cache. In experiments with high-dimensional vector search, TurboQuant consistently surpassed PQ and RaBitQ methods in recall, despite fact that they used large code books and tuning to a specific dataset.

Article, Arxiv • #AI #ML #LLM #TurboQuant #Google

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Qianfan-OCR is an end-to-end model with 4B parameters for working with documents.

Main idea is one model instead of an entire pipeline.

➡️ What it can do?

• Document parsing in one pass
Without splitting into OCR → post-processing → extraction.
The model immediately outputs a structured result.

• Tables
Correctly extracts the structure of tables, rows, and values.

• Formulas
Recognizes mathematical expressions and converts them into a readable form.

• Charts & diagrams
Understands visual data and extracts meaning from it.

• Key information extraction
Automatically retrieves key fields: sums, dates, names, etc.

Previously, this required a complex stack: OCR → layout detection → table parser → rule-based extraction.

Now, all of this replaced by one model, which does everything at once. In fact, this is a step towards systems that can understand documents just like a human.

#AI #OCR #LLM #ML

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Claude Code Game Studios: 48 AI agents as a full-fledged game studio.

A solo developer with access to Claude Code can now deploy an entire game dev studio with a creative director, department leads, and specialized specialists.

The game design approaches are based on the MDA Framework, self-determination theory, and flow state design.

➡️ What is it & how it works?

Claude Code Game Studios is a template for Claude Code that organizes an AI session into a three-level hierarchy of 48 specialized agents.

At the top are 3 directors (creative, technical, and producer) working on the Opus model.

One level below are 8 department leads on Sonnet: game designer, lead programmer, art director, sound director, narrative director, and others.

The third level is specialists on Sonnet and Haiku: from gameplay programmers and economy designers to DevOps engineers and accessibility specialists.

The agents interact according to a clear protocol: vertical task delegation from top to bottom, horizontal consultations between agents of the same level, and conflict escalation to the overall leader.

Each agent works strictly within its domain and does not modify other agents' files without explicit instructions.

In addition to the agents, the template includes 37 teams covering the entire development cycle: sprint planning, code and design reviews, asset audits, idea generation, and release preparation.

A separate category is team workflows that coordinate the work of several agents on a specific feature: combat system, narrative, interface, or sound.

In parallel, 8 hooks are attached to git events and the session lifecycle. They trigger automatically: validate commits against hardcoded values and data correctness, warn about pushes to protected branches, load the current sprint's context at startup, and record results at completion.

Another layer is 11 rules attached to project paths. They are applied when editing files in the corresponding directories and set standards for each zone of the codebase.

➜ This is not an autopilot.

The agents ask questions, propose several options with pros and cons, and show a draft. But the final decision is always human, and nothing is fixed without their approval.

➜ template works with Godot 4, Unity, and Unreal Engine 5.

For each, there is a lead agent with a set of sub-specialists: for Godot, it's GDScript, shaders, and GDExtension; for Unity, DOTS/ECS, VFX, and UI Toolkit; for Unreal, GAS, Blueprints, and Replication.

Project is open to customization: agents can be added & removed, prompts can be edited and hooks can be reconfigured.

GitHub: License: MIT License: #AI #ML #Vibecoding #Gamedev #Claude

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Hyperagents

An AI system that not only improves task solving, but also enhances the process of self-improvement itself.

➡️ What is it?

Darwin Gödel Machine (DGM) demonstrated that open self-improvement is possible through the iterative generation and evaluation of improved agents, but it relies on a key assumption: that improvements in task solving (for example, in programming) lead to improvements in the process of self-improvement itself. This correspondence persists in programming, where both evaluation and modification are expressed in the same subject area, but it is generally violated in other cases. As a result, previous systems remain limited to fixed, manually defined meta-level procedures that do not evolve on their own.

They introduce Hyperagents - self-referential agents capable of modifying both their behavior in task solving and the process that generates future improvements. This enables what they call metacognitive self-modification: learning not only to perform tasks better, but also to become better at the process of improvement itself.

This approach is implemented in the form of DGM-Hyperagents (DGM-H) - an extension of DGM, in which both the behavior of task solving and the self-improvement procedure are editable and subject to evolution. In various domains (programming, scientific paper review, design of reward functions in robotics, and evaluation of mathematical Olympiad tasks), hyperagents ensure continuous quality improvement over time and outperform basic approaches without self-improvement or open evolution, as well as previous self-improving systems (including DGM). DGM-H also improves the process of generating new agents itself (for example, through persistent memory and performance tracking), and these meta-level improvements are transferable between domains and accumulate from run to run.

Full article 🆗

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Naive RAG vs Agentic RAG - an explanation at a glance

➡️ What are the well-known problems of naive RAG that Agentic RAG solves and how?

- It performs one retrieval and one generation. If the context is irrelevant, it cannot perform a second search.
- It handles all requests the same way. A simple lookup and a complex multi-hop task go through the same retrieve-then-generate pipeline.
- There is no validation. The system blindly trusts what the retriever returns.

Agentic RAG solves this by implementing decision loops at each stage.

Steps 1-2) A query rewriting agent reformulates the original query. This is not just about correcting typos - it optimizes the query for retrieval: clarifying vague formulations, decomposing complex queries into sub-queries, and expanding abbreviations.

Steps 3-5) A routing agent decides whether external context is needed at all. If not, the retrieval is skipped. If yes, a source selector chooses the best backend for the specific type of query.

Steps 6-7) The source selector directs the query to the most suitable source: vector DB for semantic search, web search for up-to-date information, or structured APIs for tabular data. The obtained context and the reformulated query are combined into a prompt.

Steps 8-9) The LLM generates the initial response.

Steps 10-12) A validation agent (also known as Corrective RAG) checks whether the response is relevant, grounded, and complete. If everything is ok - the response is returned. If not - the system returns to step 1 with a reformulated query.

This cycle is repeated several iterations until a satisfactory response is obtained or the system acknowledges that it cannot answer.

Why this works? each agent acts as a quality gate.

- The rewriter improves the accuracy of retrieval
- The router ensures the selection of the correct source
- The validator checks that the response is grounded

Errors at individual stages are caught and corrected, rather than silently being passed on through the pipeline.

Scheme in above media is just one of possible blueprints of Agentic RAG system. In production, Corrective RAG, Adaptive RAG, Self-RAG, and hybrid search (vector + lexical with reranking) are increasingly being combined depending on latency budget and accuracy requirements.

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Building a Vectorless RAG System

(Without Embeddings & a Vector Database)

The construction of a Vectorless, Reasoning-Based RAG system using hierarchical page indexing.

➡️ What's inside?

The document is transformed into a tree, and the LLM traverses this tree to find an answer. Without embeddings. Without searching for similarity.

This is very similar to how we search for information in real life. When we need to find something in a textbook, we don't read all the pages in a row. First, we open the table of contents, find the needed chapter, then browse the sections and go directly to the required one.

PageIndex works similarly. You pass it a document, and it builds a tree where each branch is a section, and each leaf is the text itself. When you ask a question, the LLM traverses the tree level by level to find the correct answer.

Full code: GitHub

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

One of the most powerful open-source models for computer vision SAM 3.1 released

Model understands what is happening in an image or video and is able to find objects based on a text description. You can literally write "a person in a red T-shirt" and it will find necessary people.

➡️ How it works?

It works not only with pictures, but also with videos. The object can be specified once, and then the model will track it between frames.

The key idea - open-vocabulary. The model is not limited to fixed classes, like old systems. It operates with a huge number of concepts and can find almost any object.

Another important point is that you can combine control methods: text, clicks, frames, masks. This gives much more control and accuracy.

Under the hood a new architecture, where the tasks of object search and tracking are solved separately. Due to this, the model better distinguishes similar things and works more stably on video.

The repository already has everything for getting started: ready weights, code, examples, and notebooks.

In fact, this is no longer just a tool for labeling, but a full-fledged vision engine that can be integrated into real products from video analytics to data labeling automation.

Now the model can track up to 16 objects in one pass.

With multiplexing, all objects are processed simultaneously:

• fewer unnecessary calculations
• no memory bottlenecks

RESULT: video processing speed increases by approx 2 times, from 16 to 32 FPS on a single NVIDIA H100!

On new SA-CO benchmark, which includes 270 thousand unique concepts, SAM 3 achieves 75–80% of human level.

GitHub | #AI #ML #LLM #CV #python

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Vector Index vs Vector Database

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

NVIDIA proved,
Backpropagation is not only way to create AI.

They trained models with billions of parameters without using a single gradient.

Every AI you use today relies on backpropagation.

This requires complex calculations, enormous memory costs, and huge GPU clusters.

Meanwhile, an ancient method without gradients, called evolutionary strategies (ES), was considered impossible to scale.

But now everything has changed.

NVIDIA and Oxford have just introduced EGGROLL.

Instead of generating huge full-rank matrices for each mutation, they split them into two small ones.

The AI mutates. It tests. It preserves what works. Just like biological evolution.

But now it does this with hundreds of thousands of parallel mutations at once.

The performance is now the same as that of batch inference.

They pre-train models from scratch, using only simple integers.

No backpropagation. No decimals. No gradients.

We thought the future of AI required endless clusters of precision equipment.

It turned out we just needed to evolve.

Paper

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

found a version of tSNE for GPUs.

It works 33 times faster by changing just one line of code.

Within the GPU environment:

- Delete this line → from sklearn.manifold import TSNE
- Add this line → from tsnecuda import TSNE

Now, tSNE will run on the GPU for faster processing.

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Holo3 is an SOTA-agent model for managing desktop applications.

Parisian startup H Company released a family of multimodal Holo3 models designed for managing graphical interfaces.

➡️ Why Holo3 is interesting?

➜ Both models of the family on the Mixture-of-Experts architecture

The senior Holo3-122B-A10B is only available on the platform of H Company at a price of 40 cents per million incoming and $3 per million outgoing tokens.

The younger version Holo3-35B-A3B is available on Hugging Face under the Apache 2.0 license and is also available free of charge via the Inference API with a limit of 10 PRMs. In the paid mode - $0.25/1.8 per million incoming/outgoing tokens.

➜ Holo3 was trained in a closed loop of 3 stages

First, synthetic navigation examples are generated for interfaces according to specified scenarios.

Then the data is expanded beyond the initial conditions so that the model takes into account non-standard situations.

In the final stage, all examples undergo curated selection and reinforcement learning.

For training, H Company built a generator of synthetic corporate environments, in which agents create web applications according to scenario specifications, forming verifiable tasks of varying complexity.

Based on these environments, H Corporate Benchmarks was developed - a set of 486 multi-step tasks in 4 categories: e-commerce, business software, collaboration tools, and inter-application scenarios.

The latter require coordination between several systems at once (for example, extracting prices from a PDF, comparing them with employees' budgets, and automatically sending personalized letters with approval or rejection).

➜ Tests

The flagship Holo3-122B-A10B scored 78.85% on the OSWorld-Verified benchmark - this is the best result on the leading desktop interaction test.

Holistic AI founded by Charles Cantor (Ex Stanford researcher) and Laurent Cifre (Ex Google DeepMind, contributor AlphaGo project) closed seed round of $220 million in 2024 - one of largest in Europe. investors were Eric Schmidt, Yuri Milner, Bernard Arnault, Xavier Niel, as well as Amazon, Samsung, UiPath.

Apache 2.0 Licensing: Article, Model | #AI #ML #MMLM #Holo3 #HCompany

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore

Hybrid search = the accuracy of lexical search + the "intuition" of semantic search.

With ES|QL and multi-stage retrievers, you can combine in a single query:
• lexical search,
• semantic search,
• filters, and
• reranking.

••••••••••••••••••••••••••••••••••••••
🤖 Data & ML | @DataXplore