搜索“Agent”

single task. Amp is better, faster, and cheaper when used this way. A notice will appear when you hit 20% of the context window to remind you of this. Longer threads are more expensive, both because each iteration of the agentic loop sends more and more tokens, and because requests with more than 200k tokens are roughly twice as expensive per token in Anthropic's API pricing. Note: the screenshot shows 968k tokens because the context window is composed of 968k input tokens and 32k output tokens.

匹配字段正文

Over the past 12 months, tens of thousands of developers have built and productionized industry-leading voice agents using our frontier speech models, Sonic and Ink. In serving these developers, we found that even the most sophisticated teams struggled to build great voice agents with the tools available today. Voice agents are uniquely hard to bring to production: they require maintaining natural, low-latency conversations with advanced reasoning across millions of calls. A few challenges come

匹配字段正文

leaders. Google adds AI image editing features in Slides and Vids OpenAI weighs encryption for temporary chats to strengthen privacy Salesforce partners report slower adoption of Agentforce in the enterprise Apple Intelligence features surface in the iOS 26 public beta, shaping user expectations at work Google adds AI image editing to Slides and Vids Summary: Google Workspace introduced background removal and background expansion for images in Slides and Google Vids. The weekly recap also highl

匹配字段正文

estion, and it will break it down, gather credible sources, and build a structured, reference-backed report that’s easy to follow. It’s powered by our tool-augmented Deep Research agent (in preview), but designed to feel simple, transparent, and genuinely helpful, as though you’re collaborating with a well-organized research partner. Vocal mode lets you talk to Le Chat just like you’d talk to a person — no typing needed. You can brainstorm ideas while walking, get quick answers while running err

匹配字段正文

One of our most important objectives at OpenHands is to make state-of-the-art software development agents accessible to any developer no matter what your circumstances. As an open source AI coding agent, developers often use OpenHands locally through a GUI or CLI with an API key to their LLM provider of choice. However, many OpenHands users find managing API keys for multiple LLM providers a nuisance, and low rate limits leading to a sub-optimal experience. This is why we are announcing a new Op

匹配字段正文

There's a lot of news in agents nowadays, ones that code, ones that browse the web, ones that perform customer service, ones that wash your socks. However, while there is a lot of excitement around agents, there is also a lot of disappointment. When you actually try them out, they often make simple mistakes, get stuck in loops, and don't get the job done. At OpenHands, we have been working hard at making agents that actually work, and doing it in the open so you can actually understand why they

匹配字段正文

Code is powered by four models that are state of the art in coding: Codestral for fill-in-the-middle / code autocomplete Codestral Embed for code search and retrieval Devstral for agentic coding And Mistral Medium for chat assistance Critically, customers can fine-tune or post-train the underlying models on private repositories or distill lightweight variants—capabilities that simply don’t exist in closed copilots tied to proprietary APIs. Mistral Code is proficient in 80+ programming languages

匹配字段正文

globally distributed setting, we can achieve even higher speeds: Can it be used in HPC settings? Despite the fact that PCCL does not directly optimize for high-performance local networks, PCCL remains competitive with PyTorch’s Gloo over Ethernet. Usage in our prime framework If you want to see how we use PCCL, refer to the open source prime repository, which is our production-ready open source implementation of DiLoCo and async DiLoCo. ‍ Rewriting Prime Agent in Rust On the Nature of the Swarm

匹配字段正文

It can sometimes be hard to tell whether Devin (or any other coding agent) is likely to succeed with a task. When Devin tries to tackle a task that's outside of its current capabilities, this sometimes leads to wasted time and ACUs. And, one of the biggest blockers to successful sessions is simply that initial prompts are underspecified—they don't contain enough detail for Devin to use. Today, we’re announcing some major changes in Devin 2.1, with these issues in mind. Devin now reports its conf

匹配字段正文

Yannick Schnider from the Spyre team on vLLM for pluggable scheduler design and implementation, and other contributors, including yancong for extendable quantization method design and implementation, Aviv Keshet for extendable SamplingParams. Related Posts Announcing vllm-metal: Concurrent Serving on Apple Silicon vllm-metal brings vLLM's paged, continuously batched serving stack to Apple Silicon, with flatter TTFT under concurrent agent load, batched MTP, and automatic M5 prefill acceleration.

匹配字段正文

PVCs at a fraction of the cost without sacrificing quality or latency. As a result, we’ve been able to make PVCs available self-serve. Whether you’re building virtual avatars, AI agents, or assembling a private voice library, PVCs are now easily accessible through our website or API (no Sales contact needed). Fine-tune One—or Many We built our PVCs in response to overwhelming demand from businesses that need bespoke AI voices for private voice libraries and virtual personal avatars. Our custome

匹配字段正文

waits by shifting memory-bound operations toward compute-heavy execution; Reducing Memory Footprint: Investigating techniques to minimize memory requirements during decoding, including lighter cache management and strategic re-computation; Enabling Asynchronous Execution: Designing naturally asynchronous inference protocols that can tolerate variability in node availability, latency, and bandwidth without sacrificing throughput References Rewriting Prime Agent in Rust On the Nature of the Swarm

匹配字段正文

we first shared Devin 1.0 with the world last March. Since then, we've been working hard to improve Devin based on your feedback. Today, we're excited to announce Devin 2.0: a new agent-native IDE experience for working with Devin, along with a flexible new plan starting at $20. Now you can: Spin up multiple parallel Devins, each equipped with its own interactive, cloud-based IDE. This means you can easily multitask, tackling numerous tasks concurrently and stepping in to steer when needed; Coll

匹配字段正文

A system for analyzing and intervening on agent behavior We built a system, Docent, that accelerates analysis of AI agent transcripts. It helps users automatically identify corrupted tasks, fix scaffolding issues, uncover unexpected behaviors, and understand an agent's weaknesses—in only minutes. Explore a technical preview of Docent here, and sign up to join our user community and receive early updates. Docent identifies issues in an agent's environment, like missing packages the agent invoked.

匹配字段正文

ect with custom tools, and use customized models that best understand your enterprise context. Please let us know if you’d like to get early access. Data connectors and multi-step agents (coming soon) You will soon be able to plug in le Chat to your work environment (documents, email, messaging systems, databases) with granular access control and create multi-step agents to automate the boring parts of your work. What can you do with it? With le Chat now available on mobile, you can take it anyw

匹配字段正文

implementing the speculative decoding support; Kuntai Du is implementing the prefill disaggregation and KV Cache transfer support; Simon Mo and Zhuohan Li contributed to the V1 system design Related Posts DeepSeek-V4.1-Flash on vLLM: 5x Agentic Throughput Since Day 0 Within three weeks of release, vLLM made DeepSeek-V4.1-Flash 1.9x faster at low concurrency and lifted its throughput 5x on SemiAnalysis AgentX, with SWA bounded replay, CUDA graphs, DeepSeek's new kernels, and vLLM kernel fusions.

匹配字段正文

We are releasing code to train agents that can do challenging multi-step tasks in biology research. We’ve used this code to train agents built on small open source models to accuracies exceeding frontier language models and human PhD researchers at dramatically lower cost. This marks a major milestone in FutureHouse’s journey to automate scientific discovery with AI. Read the preprint: arxiv.org/abs/2412.21154 See the Aviary environment code: github.com/Future-House/aviary See the agent code: gi

匹配字段正文

ng hard for for the past couple months: we now have an official hosted version of the OpenHands AI software developer! OpenHands is an open source project for software development agents, and now boasts the strongest AI software development agent in the world, resolving over 50% of real github issues on the standard SWE-Bench Verified benchmark. OpenHands will always remain open source and free for every developer to run locally, but at the same time, spinning up the software and development san

匹配字段正文

10× to 10B-parameter model size, 25x from the original research. This brings us a step further towards to democratized training of frontier open foundation models: from language, agents to scientific models. Our goal is to solve distributed training step-by-step to ensure AGI will be open-source, transparent, and accessible. Launch Partners and Contributors We are excited and grateful to be joined by leading open-source AI players like Hugging Face, SemiAnalysis, Arcee, Hyperbolic, Olas, Akash,

匹配字段正文

ilable to run using Ollama. 1B and 3B models are text-only models are optimized to run locally on a mobile or edge device. They can be used to build highly personalized, on-device agents. For example, a person could ask it to summarize the last ten messages they received on WhatsApp, or to summarize their schedule for the next month. The prompts and responses should feel instantaneous, and with Ollama, processing is done locally, maintaining privacy by not sending data such as messages and other

匹配字段正文