Search Results for "optimize"

DeepSeek-V3.2-Exp is an experimental release of the DeepSeek model family, intended as a stepping stone toward the next generation architecture. The key innovation in this version is DeepSeek Sparse Attention (DSA), a sparse attention mechanism that aims to optimize training and inference efficiency in long-context settings without degrading output quality. According to the authors, they aligned the training setup of V3.2-Exp with V3.1-Terminus so that benchmark results remain largely comparable, even though the internal attention mechanism changes. In public evaluations across a variety of reasoning, code, and question-answering benchmarks (e.g. ...

BitNet (bitnet.cpp) is a high-performance inference framework designed to optimize the execution of 1-bit large language models, making them more efficient for edge devices and local deployment. The framework offers significant speedups and energy reductions, achieving up to 6.17x faster performance on x86 CPUs and 70% energy savings, allowing the running of models such as the BitNet b1.58 100B with impressive efficiency.

DeepSeekMath-V2 is a large-scale open-source AI model designed specifically for advanced mathematical reasoning, theorem proving, and rigorous proof verification. It’s built by DeepSeek as a successor to their earlier math-specialist models. Unlike general-purpose LLMs that might generate plausible-looking math but sometimes hallucinate or mishandle rigorous logic, Math-V2 is engineered to not only generate solutions but also self-verify them, meaning it examines the derivations, checks...

xformers is a modular, performance-oriented library of transformer building blocks, designed to allow researchers and engineers to compose, experiment, and optimize transformer architectures more flexibly than monolithic frameworks. It abstracts components like attention layers, feedforward modules, normalization, and positional encoding, so you can mix and match or swap optimized kernels easily. One of its key goals is efficient attention: it supports dense, sparse, low-rank, and approximate attention mechanisms (e.g. ...

...It is the second phase of the Chinese LLaMA&Alpaca large model project. The Chinese LLaMA-2 base model and the Alpaca-2 instruction fine-tuning large model are open-sourced. These models expand and optimize the Chinese vocabulary on the basis of the original Llama-2, use large-scale Chinese data for incremental pre-training, and further improve the basic semantics and command understanding of Chinese. Performance improvements. The related model supports FlashAttention-2 training, supports 4K context and can be extended up to 18K+ through the NTK method.

Grok-1 is a 314-billion-parameter Mixture-of-Experts (MoE) large language model developed by xAI. Designed to optimize computational efficiency, it activates only 25% of its weights for each input token. In March 2024, xAI released Grok-1's model weights and architecture under the Apache 2.0 license, making them openly accessible to developers. The accompanying GitHub repository provides JAX example code for loading and running the model. Due to its substantial size, utilizing Grok-1 requires a machine with significant GPU memory. ...

...The model was refined through a two-stage post-training process: Supervised Fine-Tuning (SFT) to transfer knowledge from Jan-v1, followed by Reinforcement Learning with Verifiable Rewards (RLVR) to optimize reasoning, tool use, and correctness. With just 1.7B parameters, Jan-v1-edge achieves 83% accuracy on SimpleQA tasks, approaching the performance of larger models like Jan-nano-128k. Benchmark comparisons show it remains competitive or superior in areas such as EQBench and recency QA, though with slight trade-offs in instruction following and creative writing compared to similar-sized Qwen models.