This article derives DeepSeek V3's training Model FLOPs Utilization (MFU) using publicly available data, showing an MFU of roughly 37%—about a 60% improvement over V2—and provides detailed formulas, parameter settings, and a reproducible Python script.
The article introduces the Think When You Need (TWYN) method, a reinforcement‑learning approach that dynamically adapts chain‑of‑thought length, dramatically cuts redundant token generation in large language models, and maintains or improves accuracy across diverse reasoning benchmarks.
DeepSeek’s distillation technology combines data and model distillation to transfer knowledge from large teacher models to compact student models, detailing its definitions, principles, key innovations, architecture, training methods, performance gains, and challenges, especially in multimodal contexts.
The article presents a two‑step cross‑architecture distillation method that replaces the quadratic softmax attention of Transformers with a learned linear attention and then maps it onto a Mamba backbone, achieving near‑teacher performance while reducing inference cost to linear time.
The article explains what tokens are, how different models tokenize text, the role of token embeddings, positional encoding, self‑attention, KV cache, and why output tokens cost far more than input tokens, while also covering pricing differences and prompt‑caching savings across major LLM providers.
Ant’s newly released Ling‑2.6‑flash model, hidden as the anonymous “Elephant Alpha,” combines a 104B‑parameter MoE design with only 7.4B active weights per inference, achieving ten‑fold token savings, top‑tier benchmark scores and a $0.10 per‑million‑token price that dramatically cuts inference costs for developers and enterprises.
HASS aligns training and decoding contexts and objectives for speculative sampling, using harmonized objective distillation and multi-step context alignment, achieving 2.81–4.05× speedup and 8%–20% improvement over EAGLE‑2 while preserving generation quality in real-world deployments at Xiaohongshu.
This article examines generative data‑driven model distillation as a technique that not only compresses large language models but also improves their accuracy, addresses data‑privacy constraints, and reduces computational costs, offering a practical roadmap and real‑world results from a corporate AI platform.
AngelSlim introduces a full‑stack large‑model compression suite that uses quantization‑aware training to shrink a 1.8B LLM to 2‑bit precision, achieving less than 4% accuracy loss, supporting a wide range of models, speculative decoding, and providing end‑to‑end deployment instructions for MacBook M4 and server environments.
A small tech firm, tngtech, released an open‑source model fusion called DeepSeek‑R1T‑Chimera that merges R1 inference with V3‑0324 without fine‑tuning, distillation, or prompts, achieving the same intelligence as R1 while reducing token output by 40% and speeding up inference.
Shanghai AI Laboratory’s InternLM 3.0 upgrade demonstrates that a refined 4 TB token dataset can boost a large‑language model’s performance beyond that of open‑source peers trained on 18 TB, cutting training cost by over 75% while merging regular dialogue with deep reasoning capabilities.
This note surveys four open‑source reproductions of DeepSeek R1/R1‑zero reinforcement‑learning pipelines, re‑implements their training on math and logic datasets using Qwen‑based models, shows that format‑plus‑accuracy rewards improve long‑chain reasoning though stability and scaling remain challenges, and outlines future directions for large‑scale RL and business deployment.
UC Berkeley researchers found that forcing large language models to skip explicit reasoning—using a “NoThinking” mode—can achieve comparable or better accuracy with significantly fewer tokens, especially under token budget constraints, across math, coding, and theorem‑proving benchmarks.
The paper introduces ReLaX, a framework that shifts focus from token‑level entropy to the latent‑space dynamics of large models, employing Koopman operators and a Dynamic Spectral Divergence metric to quantitatively guide exploration‑exploitation balance, and demonstrates state‑of‑the‑art performance on both pure‑text and multimodal RL benchmarks.
This article reviews recent 2025 advances in long‑sequence LLM inference, covering Kimi Linear attention, DuoAttention and DeepSeek Sparse Attention, MegaKernel and MPK designs for kernel‑level efficiency, reinforcement‑learning rollout optimizations, and the Tawa deep‑learning compiler framework.
The article examines the rapid token consumption problem caused by popular LLM agents, proposes a four‑tier model hierarchy and concrete routing rules, and offers short‑term, long‑term, and budget‑friendly deployment recommendations to reduce expenses while maintaining performance.
This article systematically compares the architectures of recent large language models—including DeepSeek V3/R1, OLMo 2, Gemma 3, Mistral Small 3.1, Llama 4, Qwen 3, SmolLM 3 and Kimi 2—highlighting innovations such as MLA, MoE, post‑norm, sliding‑window attention, NoPE and optimizer choices, with diagrams and code examples to illustrate their impact on efficiency and performance.
This guide walks you through deploying Qwen3.5's 0.8B, 2B, 4B and 9B models on CPUs or modest GPUs using Unsloth's GGUF quantization, explains hardware requirements, shows how to run them with llama.cpp, llama‑server, vLLM or SGLang, and provides a free Colab fine‑tuning workflow with export options.
Qwen1.5‑MoE‑A2.7B is a 2.7 billion‑parameter Mixture‑of‑Experts model that delivers performance comparable to leading 7 billion‑parameter LLMs while cutting training cost by 75% and boosting inference speed by 1.74×, and the article details its architecture, benchmarks, efficiency analysis, and deployment steps.
LongCat-Flash-Chat, an open‑source 560‑billion‑parameter Mixture‑of‑Experts model that activates only 18.6‑31.3 B parameters per token, delivers state‑of‑the‑art performance on general, agentic, coding, and instruction‑following benchmarks while offering fast inference and efficient deployment options.
