搜索“RAG”

models. We discover that improving the expressiveness of the router substantially improves performance of the overall MoE model, as well as changing the routing dynamics and encouraging greater expert specialization. Our novel ZAYA1 router replaces the linear router with a downprojection and then performs multiple sequential MLP operations within the compressed latent space to obtain the final routing choices. Together, these architectural innovations reduce the fundamental compute and memory b

匹配字段正文

isms in our environments: Classical tests (e.g. unit tests, integration tests) for reliably validating correctness; Rubrics for code quality and approach; Agentic grading that leverages a browser-use agent to test end-to-end functionality of product features These grading mechanisms usually complement each other well and many environments use a combination of all three at once. To ensure our environments are robust to reward hacking, we developed a process we call reward hardening, where human e

匹配字段正文

perspectives. First, price versus performance. A suitable model should have good performance and be affordable, falling into the green area in the chart below. Here, we use the average score from the 10 test sets on Artificial Analysis to represent performance: Next is price versus inference speed. When deploying a model, there is often a trade-off where a slower inference speed can lead to a lower price. An ideal model should be affordable and fast. We have compared several representative model

匹配字段正文

observed that our agent trajectories were often spending 60% of their first turn just retrieving context. Context retrieval has been historically done in 2 ways: Embedding Search (RAG): once the upfront work of indexing the codebase is done, queries are fast. But:The results can be inaccurate, especially for complex queries that require to jump across the codebase multiple times (such as tracing through some execution path in a large codebase); The embeddings can even be counterproductive, as th

匹配字段正文

text–vision fusion for lossless, unified comprehension Download Ollama; Run the model Prompt the model with a message and image path(s). It is possible to use multiple images and drag and drop in images to make it easier to automatically type the file path. Examples Flower identification Prompt: What is this flower? Is it poisonous to cats? Menu understanding and translation Prompt: Show me the menu in English! JavaScript Library Install Ollama’s JavaScript library Pull the model Example non-str

匹配字段正文

a 1.5B dense model, while the 8.3B total capacity lets experts specialize (reasoning, multilingual, code, long-tail knowledge) and lifts quality. Memory & deployability. Weight storage scales with total parameters, while compute and state caches scale with the active path. In practice, quantized variants fit comfortably on high-end phones, tablets, and laptops for a wide range of applications. Inference The following video shows LFM2-8B-A1B running locally on a MacBook Pro with an Apple M2 Pro c

匹配字段正文

eractive use. The model is provided with an input waveform (4 seconds), and the time to generate the first audible sound from the model is measured. LFM2-Audio-1.5B achieved an average end-to-end latency of under 100 ms, highlighting superb efficiency, even faster than models much smaller than 1.5B parameters. Deploy and build with LFM2-Audio Thanks to its speed, quality, and multimodal understanding, LFM2-Audio unifies what used to be an entire pipeline of models. This allows developers to buil

匹配字段正文

ing Sora that it would have been if they hadn’t. If that’s not the case, we will make significant changes (and if we can’t fix it, we would discontinue offering the service). Encourage users to control their feed. You should be able to tell Sora what you want—do you want to see videos that will make you more relaxed, or more energized? Or only videos that fit a specific interest? Or only for a certain about of time? Eventually as our technology progresses, you will be should to the tell Sora wha

匹配字段正文

ng to architecture with deep contextual understanding, transforming our development velocity.Eric Wendelin Tech Lead, GenAI for Developer Productivity Claude Sonnet 4.5 reduced average vulnerability intake time for our Hai security agents by 44% while improving accuracy by 25%, helping us reduce risk for businesses with confidence.Nidhi Aggarwal Chief Product Officer Claude Sonnet 4.5 is state of the art on the most complex litigation tasks. For example, analyzing full briefing cycles and conduc

匹配字段正文

cart are building AI tools that automate workflows, track media, summarize content, and draft reports, all with natural-language interfaces rather than code. Smaller teams are leveraging third-party LLMs or agencies for similar capabilities. Why it matters: Deploying AI for internal comms can free up teams from manual tasks and shift focus to strategic storytelling and reputation management. Source: Axios Google Nano Banana (Gemini 2.5 Flash Image) enhances image generation consistency Summary:

匹配字段正文

emphasis on image understanding. Vision training data comes from a combination of large-scale open-source datasets and in-house synthetic vision datasets, selected to balance coverage across diverse tasks. Overall, LFM2-VL is trained on the order of 100 billion multimodal tokens. Evaluation Benchmarks We evaluate LFM2-VL on several public vision-language benchmarks. The model shows excellent performance in high-resolution image understanding and multimodal instruction following, while maintaini

匹配字段正文

RO GPUs NVIDIA and Ollama are advancing their partnership to boost model performance on NVIDIA GeForce RTX and RTX PRO GPUs. This collaboration enables users on RTX-powered PCs to accurately leverage the capabilities of OpenAI’s gpt-oss model. We will continue to collaborate and enhance Ollama. In the future, we will publish an in-depth engineering post on the model. The model can be downloaded directly in Ollama’s new app or via the terminal: OpenAI launch blog OpenAI model card NVIDIA RTX blog

匹配字段正文

ten need to adapt models to their own codebases and development conventions. Without access to model weights, pos-training workflows, or extensibility, teams are locked out of leveraging the best of their codebases. Fragmented architecture: Agents, embeddings, completions, and plugins are frequently decoupled across vendors—leading to integration drift, inconsistent context handling, and operational overhead. Moreover, coding copilots are not well-integrated into full enterprise platforms, such

匹配字段正文

r less than half the price. To assess Voxtral’s transcription capabilities, we evaluate it on a range of English and multilingual benchmarks. For each task, we report the macro-average word error rate (lower is better) across languages. For English, we report a short-form ( 30-seconds) average. Voxtral comprehensively outperforms Whisper large-v3, the current leading open-source Speech Transcription model. It beats GPT-4o mini Transcribe and Gemini 2.5 Flash across all tasks, and achieves state-

匹配字段正文

scores despite being smaller in size. Inference We exported LFM2 to multiple inference frameworks to accommodate a variety of deployment scenarios. For on-device inference, we leveraged both the PyTorch ecosystem via ExecuTorch and the open-source llama.cpp library. The LFM2 models were evaluated using each platform's recommended quantization schemes - 8da4w for ExecuTorch and Q4 0 for llama.cpp - and benchmarked against existing models available in those ecosystems. Target hardware included the

匹配字段正文

mine, modify and build upon its architecture and reasoning processes to further accelerate the emergence of thinking language models. Our earlier open models have already been leveraged by the community for exciting projects like ether0 and DeepHermes 3 . Magistral is fine-tuned for multi-step logic, improving interpretability and providing a traceable thought process in the user’s language, unlike general-purpose models. We aim to iterate the model quickly starting with this release. Expect the

匹配字段正文

t One of our core improvements on standard Whisper is dynamic chunking. Standard Whisper performs best on full 30-second chunks. But conversational AI deals in much smaller, more fragmented audio segments. We’ve modified Whisper to handle variable-length chunks that end at semantically meaningful points. That means fewer errors and less hallucination, especially during silence or audio gaps. To ensure the Ink-Whisper actually works better in the wild, we created a suite of evaluation datasets th

匹配字段正文

, Tamil, Gujarati, Urdu, Kannada, Odia, Malayalam, Punjabi, Assamese, Maithili, Santali, Kashmiri, Nepali, Sindhi, Dogri, Konkani, Manipuri (Meitei), Bodo, Sanskrit. It supports paragraph-level translation for 22 languages, and supports translating diverse structured content for 15 languages. In human evaluation by language experts, Sarvam-Translate is identified to be significantly better than much larger models such as Gemma3-27B-IT, Llama4 Scout, and Llama-3.1-405B-FP8. Further, in automated

匹配字段正文

e embedding model. Codestral Embed can output embeddings with different dimensions and precisions, and the figure below illustrates the trade-offs between retrieval quality and storage costs. Codestral Embed with dimension 256 and int8 precision still performs better than any model from our competitors. The dimensions of our embeddings are ordered by relevance. For any integer target dimension n, you can choose to keep the first n dimensions for a smooth trade-off between quality and cost. Below

匹配字段正文

d communication-free, so more complex RL environments that devote more FLOPs to inference are a natural fit for low-communication training; Tool Calls & Multi-Turn RL To fully leverage inference-time compute for scientific and research progress, we need to give models built-in tools in its reasoning chain—web search, Python interpreters, etc; Crowdsourcing RL Tasks & Environments We believe open-source has a unique advantage here. Distributed RL is still in its early days, and with the right com

匹配字段正文