Search Results for "rope"

GLM-4 is a family of open models from ZhipuAI that spans base, chat, and reasoning variants at both 32B and 9B scales, with long-context support and practical local-deployment options. The GLM-4-32B-0414 models are trained on ~15T high-quality data (including substantial synthetic reasoning data), then post-trained with preference alignment, rejection sampling, and reinforcement learning to improve instruction following, coding, function calling, and agent-style behaviors. The...

Qwen2.5-Omni is an end-to-end multimodal flagship model in the Qwen series by Alibaba Cloud, designed to process multiple modalities (text, images, audio, video) and generate responses both as text and natural speech in streaming real-time. It supports “Thinker-Talker” architecture, and introduces innovations for aligning modalities over time (for example synchronizing video/audio), robust speech generation, and low-VRAM/quantized versions to make usage more accessible. It holds...

...The model is trained from scratch, reportedly on a vast multilingual + code + reasoning dataset, and competes with other open or open-weight models. The architecture mirrors established decoder-only transformer families: pre-norm structure, rotational embeddings (RoPE), grouped query attention (GQA), and mixing in languages and tasks. It supports both “Base” (foundation model) and “Chat” (instruction / conversation tuned) variants.

...Recommended usage involves prompts starting with <think>\n, non-greedy sampling strategies, and support for standardized outputs on math and multiple-choice tasks. For long input handling, it supports YaRN (Yet another RoPE Namer) for context scaling.

...It demonstrates improved performance in coding, mathematics, and multilingual understanding across over 29 languages. Qwen2.5-14B-Instruct is built on a transformer backbone with RoPE, SwiGLU, RMSNorm, and attention QKV bias. It’s resilient to varied prompt styles and is especially effective for JSON and tabular data generation. The model is instruction-tuned and supports chat templating, making it ideal for chatbot and assistant use cases.

...With 80B total parameters and 3B activated at a time, it leverages hybrid attention (Gated DeltaNet + Gated Attention) and a high-sparsity Mixture-of-Experts architecture to achieve exceptional efficiency. The model natively supports a context length of 262K tokens and can be extended up to 1 million tokens using RoPE scaling (YaRN), making it highly capable for processing large documents and extended conversations. Multi-Token Prediction (MTP) boosts both training and inference, while stability optimizations such as weight-decayed and zero-centered layernorm ensure robustness. Benchmarks show it performs comparably to larger models like Qwen3-235B on reasoning, coding, multilingual, and alignment tasks while requiring only a fraction of the training cost.

Devstral Small 2 is a compact agentic language model designed for software engineering workflows, excelling at tool usage, codebase exploration, and multi-file editing. With 24B parameters and FP8 instruct tuning, it delivers strong instruction following while remaining lightweight enough for local and on-device deployment. The model achieves competitive performance on SWE-bench, validating its effectiveness for real-world coding and automation tasks. It introduces vision capabilities,...