Tencent Hunyuan has open sourced HPC-Ops, a production grade operator library for large language model inference architecture devices. HPC-Ops focuses on low level CUDA kernels for core operators such as Attention, Grouped GEMM, and Fused MoE, and exposes them through a compact-C and Python API for integration into existing inference stacks. HPC-Ops runs in large scale internal services. In those deployments it delivers about 30 percent queries per minute improvement for Tencent-HY models and about 17 percent improvement for DeepSeek models on mainstream inference cards. These gains are reported at the service level, so they reflect the cumulative effect of…

Moonshot AI has released Kimi K2.5 as an open source visual agentic intelligence model. It combines a large Mixture of Experts language backbone, a native vision encoder, and a parallel multi agent system called Agent Swarm. The model targets coding, multimodal reasoning, and deep web research with strong benchmark results on agentic, vision, and coding suites. Model Architecture and Training Kimi K2.5 is a Mixture of Experts model with 1T total parameters and about 32B activated parameters per token. The network has 61 layers. It uses 384 experts, with 8 experts selected per token plus 1 shared expert. The attention…

In this tutorial, we implement Tree-KG, an advanced hierarchical knowledge graph system that goes beyond traditional retrieval-augmented generation by combining semantic embeddings with explicit graph structure. We show how we can organize knowledge in a tree-like hierarchy that mirrors how humans learn, from broad domains to fine-grained concepts, and then reason across this structure using controlled multi-hop exploration. By building the graph from scratch, enriching nodes with embeddings, and designing a reasoning agent that navigates ancestors, descendants, and related concepts, we demonstrate how we can achieve contextual navigation and explainable reasoning rather than flat, chunk-based retrieval. Check out the FULL CODES…

Data science agents should inspect datasets, design workflows, run code, and return verifiable answers, not just autocomplete Pandas code. DSGym, introduced by researchers from Stanford University, Together AI, Duke University, and Harvard University, is a framework that evaluates and trains such agents across more than 1,000 data science challenges with expert curated ground truth and a consistent post training pipeline. https://arxiv.org/pdf/2601.16344 Why existing benchmarks fall short? The research team first probe existing benchmarks that claim to test data aware agents. When data files are hidden, models still retain high accuracy. On QRData the average drop is 40.5 percent, on DAEval…

@tool def sql_investigate(query: str) -> dict: try: df = con.execute(query).df() head = df.head(30) return { "rows": int(len(df)), "columns": list(df.columns), "preview": head.to_dict(orient="records") } except Exception as e: return {"error": str(e)} @tool def log_pattern_scan(window_start_iso: str, window_end_iso: str, top_k: int = 8) -> dict: ws = pd.to_datetime(window_start_iso) we = pd.to_datetime(window_end_iso) df = logs_df[(logs_df["ts"] >= ws) & (logs_df["ts"] <= we)].copy() if df.empty: return {"rows": 0, "top_error_kinds": [], "top_services": [], "top_endpoints": []} df["error_kind_norm"] = df["error_kind"].fillna("").replace("", "NONE") err = df[df["level"].isin(["WARN","ERROR"])].copy() top_err = err["error_kind_norm"].value_counts().head(int(top_k)).to_dict() top_svc = err["service"].value_counts().head(int(top_k)).to_dict() top_ep = err["endpoint"].value_counts().head(int(top_k)).to_dict() by_region = err.groupby("region").size().sort_values(ascending=False).head(int(top_k)).to_dict() p95_latency = float(np.percentile(df["latency_ms"].values, 95)) return { "rows": int(len(df)), "warn_error_rows": int(len(err)), "p95_latency_ms": p95_latency, "top_error_kinds": top_err,…

For decades, predicting the weather has been the exclusive domain of massive government supercomputers running complex physics-based equations. NVIDIA has shattered that barrier with the release of the Earth-2 family of open models and tools for AI weather and climate prediction accessible to virtually anyone, from tech startups to national meteorological agencies. In a move that democratizes climate science, NVIDIA unveiled 3 groundbreaking new models powered by novel architectures: Atlas, StormScope, and HealDA. These tools promise to accelerate forecasting speeds by orders of magnitude while delivering accuracy that rivals or exceeds traditional methods. The Democratization of Weather Intelligence Historically, running…

Clawdbot is an open source personal AI assistant that you run on your own hardware. It connects large language models from providers such as Anthropic and OpenAI to real tools such as messaging apps, files, shell, browser and smart home devices, while keeping the orchestration layer under your control. The interesting part is not that Clawdbot chats. It is that the project ships a concrete architecture for local first agents, and a typed workflow engine called Lobster that turns model calls into deterministic pipelines. Architecture: Gateway, Nodes and Skills At the center of Clawdbot is the Gateway process. The Gateway…

We initiate this tutorial by configuring a high-performance evaluation environment, specifically focused on integrating the DeepEval framework to bring unit-testing rigor to our LLM applications. By bridging the gap between raw retrieval and final generation, we implement a system that treats model outputs as testable code and uses LLM-as-a-judge metrics to quantify performance. We move beyond manual inspection by building a structured pipeline in which every query, retrieved context, and generated response is validated against rigorous academic-standard metrics. Check out the FULL CODES here. import sys, os, textwrap, json, math, re from getpass import getpass print("🔧 Hardening environment (prevents common Colab/py3.12…

StepFun has introduced Step-DeepResearch, a 32B parameter end to end deep research agent that aims to turn web search into actual research workflows with long horizon reasoning, tool use and structured reporting. The model is built on Qwen2.5 32B-Base and is trained to act as a single agent that plans, explores sources, verifies evidence and writes reports with citations, while keeping inference cost low. From Search to Deep Research Most existing web agents are tuned for multi-hop question-answering benchmarks. They try to match ground truth answers for short questions. This is closer to targeted retrieval than to real research. Deep…

def visualize_results(df, priority_scores, feature_importance): fig, axes = plt.subplots(2, 3, figsize=(18, 10)) fig.suptitle('Vulnerability Scanner - ML Analysis Dashboard', fontsize=16, fontweight="bold") axes[0, 0].hist(priority_scores, bins=30, color="crimson", alpha=0.7, edgecolor="black") axes[0, 0].set_xlabel('Priority Score') axes[0, 0].set_ylabel('Frequency') axes[0, 0].set_title('Priority Score Distribution') axes[0, 0].axvline(np.percentile(priority_scores, 75), color="orange", linestyle="--", label="75th percentile") axes[0, 0].legend() axes[0, 1].scatter(df['cvss_score'], priority_scores, alpha=0.6, c=priority_scores, cmap='RdYlGn_r', s=50) axes[0, 1].set_xlabel('CVSS Score') axes[0, 1].set_ylabel('ML Priority Score') axes[0, 1].set_title('CVSS vs ML Priority') axes[0, 1].plot([0, 10], [0, 1], 'k--', alpha=0.3) severity_counts = df['severity'].value_counts() colors = {'CRITICAL': 'darkred', 'HIGH': 'red', 'MEDIUM': 'orange', 'LOW': 'yellow'} axes[0, 2].bar(severity_counts.index, severity_counts.values, color=[colors.get(s, 'gray') for s in severity_counts.index]) axes[0, 2].set_xlabel('Severity') axes[0, 2].set_ylabel('Count') axes[0, 2].set_title('Severity Distribution') axes[0, 2].tick_params(axis="x",…