Blogs
2026
How Good Are GLM5.2 and Kimi-K3: Create Your Own Benchmark and Revisit Your Training Recipe
A systematic evaluation of three open-source model families — Qwen3.5 (35B/122B/397B), GLM5.2-744B-A40B, and Kimi-K3 (2.8T/104B active) — on a self-created vertical-domain agentic multi-turn benchmark. Accuracy scales monotonically with active parameters across all three families. Kimi-K3 leads (0.492 pass@1, 0.632 pass@5) — but the headline finding is the turn gap: K3 matches GLM5.2’s per-answer accuracy at 3.3× fewer turns, and the two models fail differently (confident error vs step exhaustion). We extend our training recipe insights on PPO vs GRPO/DAPO exploration behavior, scale-dependent catastrophic forgetting, and serving-config pitfalls for self-hosting.
Designing Self-evolving Coding Agents for Terminal-Bench 2.0: A Systematic Exploration
We present a systematic exploration of designing self-evolving coding agents for Terminal-Bench 2.0. Starting from a vanilla ReAct agent powered by Claude Opus 4.6 (65.91%), we conduct large-scale controlled experiments across six groups — baselines, ceiling analysis, harness component exploration, skill evolution strategies, and self-verification mechanisms. Our best self-evolution algorithm achieves 76.47%, a +10.56pp improvement over the baseline, ranked top #7 in the Terminal-Bench 2.0 leaderboard.
Based on DeepSeek’s 13 papers, this blog post traces DeepSeek’s complete technical evolution from Jan 2024 to Jan 2026 across two major arcs: 1) the base model lineage (DeepSeek LLM → MoE → V2 → V3 → V3.2) and 2) the reasoning model lineage (Coder → Math → Coder-V2 → Prover → Prover-V1.5 → R1 → Prover-V2 → Math-V2). What emerges is a story of incremental compounding — where each paper’s contribution, however modest in isolation, becomes a critical building block for what follows.
2025
Building and Aligning LLM using User Data and RL
This blog discusses the challenges and solutions in building and aligning large language models (LLMs) using user data and reinforcement learning (RL). It walks through an end-to-end alignment stack for an open-ended dialogue model: from low-level training stability in Mixture-of-Experts (MoE) to FP8 serving, supervised fine-tuning (SFT) on user chats, and finally reinforcement learning (RL) driven by engagement and retention signals.
Awesome Cursor Training Tutorial
A comprehensive guide to understanding how Cursor scales Reinforcement Learning (RL) for AI-powered code assistance from Ray Summit 2025. Cursor uses a sophisticated RL pipeline that trains AI models to become better coding assistants by learning from real-world interactions in a production-like environment. The system consists of three main components working in harmony: (1) Trainer: Handles model training with custom optimizations; (2) Inference: Manages model deployment and load balancing; (3) Environment: Provides a production-grade agent server for realistic training.
The Hidden Language of AI: How Chat Templates Reveal the Evolution of LLMs
The blog explains that “chat templates”—the hidden formatting that turns role-tagged messages into the exact token strings models were trained on—have quietly driven LLMs’ evolution from plain text completion to multi-turn assistants to tool-enabled agents. It traces the shift from early prompt hacks to role labels and system prompts (which enforce persona and rules), then shows how different ecosystems (OpenAI, Llama, Mistral, Qwen) adopted incompatible templates, prompting Hugging Face’s Jinja-based standardization. We also provide practical guides that using the wrong template degrades quality, highlights the token-economics of template design, and forecasts templates expanding to handle tool use, multi-agent threads, million-token contexts, and richer modalities.