搜索“RAG”

obal self-attention layers in the future. at least for language modeling Hybrid models that combine the general memory-like nature of linear layers with the exact database-like storage of self-attention’s KV cache have been shown empirically to outperform pure models while enabling significant memory and compute savings 10, and we do find here that the combination of linear layers with self-attention enables better retrieval compared to a vanilla Transformer. However, we highlight that the exact

匹配字段正文

rather than traditional softmax gating, which improves expert load balancing and reduces routing collapse during training. An expert-bias term stabilizes routing dynamics and encourages more uniform expert utilization across training steps. We observed that the 105B model achieved benchmark superiority over the 30B remarkably early in training, suggesting efficient scaling behavior. Supervised Finetuning During supervised fine-tuning, the model is trained on a large corpus of high-quality prompt

匹配字段正文

s,; The inference engines sustain s roll output tokens/sec/GPU at saturation (measured end-to-end, including prefill),; r denotes the output-to-input token ratio,; L out is the average number of output tokens per trajectory,; The trainer runs at s train tokens/sec/GPU for the update workload One practical detail is that we parameterize inference cost using (1) the measured output tokens/sec on the inference engines, and (2) the average output/input token ratio. This works well in our setup becau

匹配字段正文

reasoning traces. We chose this route because it was faster to post-train an instruct version, and instruct models tend to be more popular than thinking variants. Below we show average benchmark scores across the LFM2 family, from the 350M dense model up to the 24B MoE. Across benchmarks including GPQA Diamond, MMLU-Pro, IFEval, IFBench, GSM8K, and MATH-500, quality improves log-linearly as we scale from 350M to 24B total parameters. This near-100x parameter range confirms that the LFM2 hybrid a

匹配字段正文

p install Python package: pip install zuna Organizations and researchers interested in collaborating with Zyphra to improve future versions of ZUNA for specific use cases are encouraged to reach out to bci@zyphra.com. For more information please reference the technical paper and the Zyphra blog post or visit www.zyphra.com. About Zyphra Zyphra is an open-source superintelligence company based in San Francisco, CA on a mission to build human-aligned AI that helps individuals and organizations rea

匹配字段正文

ng the same discard-all strategy as DeepSeek-V3.2 and Kimi K2.5 achieved 78.6. WideSearch: we use a 256k context window without any context management. MMLU-ProX: we report the averaged accuracy on 29 languages. WMT24++: a harder subset of WMT24 after difficulty labeling and rebalancing; we report the averaged scores on 55 languages using XCOMET-XXL. MAXIFE: we report the accuracy on English + multilingual original prompts (totally 23 settings). Empty cells (--) indicate scores not yet available

匹配字段正文

page demos. To evaluate these capabilities, we also upgraded the VIBE benchmark to a more complex and challenging Pro version, significantly increasing task complexity, domain coverage, and evaluation accuracy. Overall, M2.5 performs on par with Opus 4.5. On Droid: 79.7(M2.5) 78.9(Opus 4.6); On OpenCode: 76.1(M2.5) 75.9(Opus 4.6) Search and Tool calling Effective tool calling and search are prerequisites for a model's ability to autonomously handle more complex tasks. In evaluations on benchmark

匹配字段正文

ining detailed descriptive prompts is actually quite straightforward. For example, given this simple input: 帮我生成一个手绘风格的杭州两日禅意人文之旅双语海报 We can feed it into an LLM for rewriting, leveraging its world knowledge to produce a richly detailed prompt like this: 这是一幅中国风手绘风格的杭州两日禅意人文之旅行程导览双语海报，整体采用淡雅米黄色仿古宣纸背景，四角饰有传统回纹边框；画面中央以一条飘逸的云纹卷轴丝带贯穿连接两天行程，上方大标题为“杭州·两日禅意人文之旅”（“Hangzhou: A Two-Day Journey of Zen, Culture, and Humanity”），副标题为“祈福·山水·寻梦”（“Prayer · Landscape · Dream-Seeking”）；左侧为“第一天：灵山祈福，登高求财”（“Day 1: Pr

匹配字段正文

ese intelligence gains do not come at the cost of safety. On our automated behavioral audit, Opus 4.6 showed a low rate of misaligned behaviors such as deception, sycophancy, encouragement of user delusions, and cooperation with misuse. Overall, it is just as well-aligned as its predecessor, Claude Opus 4.5, which was our most-aligned frontier model to date. Opus 4.6 also shows the lowest rate of over-refusals—where the model fails to answer benign queries—of any recent Claude model. The overall

匹配字段正文

al at powering our AI Teammates. It's also a fantastic coding model – its ability to navigate a large codebase and identify the right changes to make is state of the art.Amritansh Raghav Interim CTO, Asana Claude Opus 4.6 reasons through complex problems at a level we haven't seen before. It considers edge cases that other models miss and consistently lands on more elegant, well-considered solutions. We're particularly impressed with Opus 4.6 in Devin Review, where it's increased our bug catchin

匹配字段正文

n edge devices, ensuring privacy and security for sensitive deployments. We’re releasing the model weights under Apache 2.0 on the Hugging Face Hub. Voxtral Mini Transcribe V2. Average diarization error rate (lower is better) across five English benchmarks (Switchboard, CallHome, AMI-IHM, AMI-SDM, SBCSAE) and the TalkBank multilingual benchmark (German, Spanish, English, Chinese, Japanese). Average word error rate (lower is better) across the top-10 languages in the FLEURS transcription benchmar

匹配字段正文

QWEN CHAT API DISCORD Introduction We present Qwen3-Max-Thinking , our latest flagship reasoning model. By scaling up model parameters and leveraging substantial computational resources for reinforcement learning, Qwen3-Max-Thinking achieves significant performance improvements across multiple dimensions, including factual knowledge, complex reasoning, instruction following, alignment with human preferences, and agent capabilities. On 19 established benchmarks, it demonstrates performance compar

匹配字段正文

.2 billion parameter model trained specifically for reasoning tasks. It generates thinking traces before producing answers, working through problems systematically. This model leverages the unique inference speed of LFMs to produce higher-quality answers. Qwen3-1.7B (thinking mode) Compared to LFM2.5-1.2B-Instruct, three capabilities jump dramatically: math reasoning (63→88 on MATH-500), instruction following (61→69 on Multi-IF), and tool use (49→57 on BFCLv3). LFM2.5-1.2B-Thinking matches or ex

匹配字段正文

ns, as a single token like “New York.” This is called a Superword Token, accounting for approximately 20% of the K-EXAONE lexicon. As a result, K-EXAONE's tokenizer achieved an average efficiency improvement of approximately 30% compared to the existing EXAONE, measured in bytes per token. [Figure 2] Improvement of tokenizer efficiency 2. Learning methods to achieve global-level performance 2.1 Pre-learning The pre-learning of K-EXAONE consists of three stages where the model sequentially learns

匹配字段正文

l tasks such as image-text retrieval, video-text matching, visual question answering (VQA), and multimodal content clustering. Unified Representation Learning (Embedding) : By leveraging the Qwen3-VL architecture, the Embedding model generates semantically rich vectors that capture both visual and textual information in a shared space. This facilitates efficient similarity computation and retrieval across different modalities. High-Precision Reranking (Reranker) : We also introduce the Qwen3-VL-

匹配字段正文

inference. These capabilities unlock new deployment scenarios across various devices, including vehicles, mobile devices, laptops, IoT devices, and embedded systems. Qualcomm Snapdragon® X Elite Qualcomm Dragonwing IQ9 (IQ-9075) (IoT) The LFM2.5 family is available now on Hugging Face, LEAP, and our playground: LFM2.5-1.2B-Base: Hugging Face; LFM2.5-1.2B-Instruct: Hugging Face, LEAP, Playground; LFM2.5-1.2B-JP: Hugging Face, LEAP; LFM2.5-VL-1.6B: Hugging Face, LEAP, Playground, Demo; LFM2.5-Audi

匹配字段正文

his suite encompasses five core subsets: Web, Simulation, Android, iOS, and Backend. MiniMax-M2.1 delivers outstanding performance on the VIBE aggregate benchmark, achieving an average score of 88.6—demonstrating robust full-stack development capabilities. MiniMax M2.1 built a 3D Dreamy Christmas Tree based on React Three Fiber and InstancedMesh, successfully rendering over 7,000 instances. It supports gesture interaction and complex particle animation, demonstrating advanced 3D rendering capabi

匹配字段正文

languages—Chinese, English, German, Italian, Portuguese, Spanish, Japanese, Korean, French, and Russian—based on the cloned voice. On the MiniMax TTS Multilingual Test Set, its average word error rate (WER) is consistently better than MiniMax, ElevenLabs, and GPT-4o-Audio-Preview. High Expressiveness ：Qwen3-TTS-VD-Flash and Qwen3-TTS-VC-Flash offer highly expressive, humanlike voices that can stably and reliably produce speech closely aligned with the input text, automatically adjusting tone and

匹配字段正文

the-art accuracy, outperforming both enterprise document processing solutions as well as AI-native OCR solutions Now powers Document AI Playground in Mistral AI Studio , a simple drag-and-drop interface for parsing PDFs/images into clean text or structured JSON Major upgrade over Mistral OCR 2 in forms, handwritten content, low-quality scans, and tables Mistral OCR 3 is designed to extract text and embedded images from a wide range of documents with exceptional fidelity. It supports markdown out

匹配字段正文

tier models. Our complete recipe — from the model weights and training frameworks, to our datasets, RL environments, and evaluations — has been open-sourced, with the goal of encouraging more open research on large scale reinforcement learning. INTELLECT-3 is trained on the same software and infrastructure that we’re open-sourcing and making available on our platform at Prime Intellect, giving everyone the tools to post-train their own state-of-the-art models, and moving us towards a future wher

匹配字段正文