搜索“Agent”

T-6 Astra, and 28x below Claude Opus 5. The new DeepSeek-V4.1-Flash model fundamentally changes the unit economics of autonomous software engineering tasks. Running high-iteration agent workflows, multi-turn bug sweeps, or large-scale test generation on expensive closed models like GPT-6 Astra and Opus 5 rapidly drains engineering budgets. At $0.43 per task, frontier-grade SWE agents become viable as continuous, background infrastructure. With DeepSeek V4.1 Flash, you can now run 15 autonomous c

匹配字段正文

DeepSeek released V4.1-Flash this week: a 552B-parameter multimodal MoE model with a Causal Encoder-Decoder architecture built for coding agents. It's available now as a Baseten Model API, with improved KV cache efficiency and asymmetric prefill/decode compute over its V4 predecessors. This week, DeepSeek released the open weights for DeepSeek-V4.1-Flash on HuggingFace , and the model is already available on Baseten Model APIs (with Loops support coming soon). Key specifications include: 552B to

匹配字段正文

.60, compared with 81.56 for Qwen 3.5 397B A17B, 76.50 for GLM 5.2 FP8, 81.37 for DeepL NextGen, 79.46 for Gemma 4 31B (on), and 68.20 for Google Translate. North Small Translate (Agentic) — which can find errors and fix errors in translation — scores even higher, at 84.36.² North Small Translate performs strongly across 32 high-resource languages and 18 additional languages. North Small Translate is the most consistent performer across the full spread of regions, without the sharp regional drop

匹配字段正文

ice, and comes within a few points of GPT-6 Astra at a quarter of the cost. SWE-2 is post-trained from Kimi K33, a 2.8T-parameter model that had already undergone extensive RL for agentic coding. As with SWE-1.7, our RL still finds substantial headroom, adding 5–6 points on many benchmarks and shifting K3’s entire cost–performance frontier. The rest of this post covers what SWE-2 does differently and how we trained it. We begin with SWE-2’s behavior, focusing on the characteristics that make it

匹配字段正文

ex, multi-step tasks? The Fugu Journey 🐡 In just a few months, Sakana Fugu has evolved from a beta thesis into an enterprise-grade orchestration engine: April (Beta): Proved multi-agent orchestration works as a unified foundation model; June (General Availability & Fugu Ultra v1): Highlighted that an orchestration layer can match closed frontier models on hard benchmarks; July (Fugu-Cyber & Claude Code Interface): Showed that orchestration can specialize in real-world domain workflows like cyber

匹配字段正文

d world didn’t end after a few seconds—but kept running for hours, responded to every move, let you attack, cast spells, shoot, or summon storms, and continued evolving through AI agents? That’s LingBot-World 2.0: our open-source real-time interactive world model, scaling to 720p/60fps in our full setup. Today, we’re open-sourcing 3 more models: 🔹 LingBot-World 2.0 Small (1.3B) 🔹 LingBot-World 2.0 Bidirectional 🔹 LingBot-World 2.0 Causal Pretrain And now, with our 1.3B Small Model, that world ca

匹配字段正文

eal test comes when users get access — and with Fable 5.1 here at the same price, the frontier now has a pretty clear head-to-head to compare against. The Rundown: Startup School: Agent Builder is a 4-part live series designed specifically for founders and lead engineers. Kicking off on Sept. 15, you’ll master the Gemini Enterprise Agent Platform and Google Antigravity 2.0 to transition your AI from rapid prototyping to enterprise-grade production. In this intensive series, you will: Code agent

匹配字段正文

计口径；一项真实的金融任务，则可能同时包含信息检索、信源核验、数据计算、逻辑分析和结果交付。 当大模型开始进入真实金融工作流，我们需要回答两个问题： 模型能否处理复杂、连续的金融任务？ 我们又该如何判断它找到的信息是否可靠、计算是否正确、答案是否有据可查？ 今天，我们正式开源面向真实金融工作流的 Ling-3.0-flash-Fin，并同步开放面向金融搜索 Agent 的评测基准 FinFIRST（Financial Information Retrieval, Sourcing and Traceability）。 Ling-3.0-flash-Fin 将金融模型能力开放给更多开发者，使模型能够进入机构自己的数据、工具和业务环境；FinFIRST 则提供一套更贴近真实金融需求的搜索任务、标准与评测方法，用于检验模型能否在得到答案的同时，给出可靠信源、清晰口径和可复核的推导过程。该测评基准由蚂蚁集团构建并开源，中金公司投行团队在构建过程中提供专业支持。 01 Ling-3.0-flash-Fin 开源：从模型可用，走向模型开放 Ling-3.0-flash-Fin（124B A5.1B

匹配字段正文

instead of following a fixed clip or script, so sessions have no preset length. This creates a foundation for interactive entertainment, virtual characters, robotics and embodied-agent simulation and generative design and interfaces. Playing the survivor. A desert world played in first person: the survivor moves, jabs the spear, drinks from a canteen and breaks into a run. Directing the scene. The same world from the director's seat: prompts addressed to the campfire and the weather flare the f

匹配字段正文

We’re excited to release Muse Spark 1.3 with improved performance on agentic and coding tasks, and a focus on real-world usability. Key capabilities: → Sustains longer-horizon work across multiple workflows in a single thread → More actively collaborates with users: it asks clarifying questions, flags when it's stuck, confirms before consequential actions → Better calibrated on its own limits instead of hallucinating outcomes → 20% fewer tool calls and 25% fewer tokens vs. Muse Spark 1.2 in inte

匹配字段正文

prediction that AI erases jobs in the ‘tens of millions’ in the next decade. He points to the many security incidents from the past month, including OAI’s hack that saw over 1,000 agents coordinating online unnoticed for weeks. Sanders calls for an international global pause, pushing for the U.S. and China to strike a “Cold War” style AI deal at their meeting in Washington later this month. Why it matters: Sanders has never been shy about his fears (he even aired them to Claude's voice mode a fe

匹配字段正文

大幅提升，往往接近成本更高的前沿模型的性能。 在 DeepSWE v1.1（长周期软件工程）上，3.8 Flash 在端到端自主解决复杂工程问题方面优于大多数更大的前沿模型，而成本仅为其一小部分。 此外，3.8 Flash 在专业知识领域展现出关键企业自主性所需的可靠性。在需要高级分析和报告的定量与专业领域，3.8 Flash 在 Vals Finance Agent V2 和 Harvey 的法律智能体基准等基准测试中优于 3.7 Flash 和其他前沿模型。3.8 Flash 在 HLE-Verified 上还取得了 54.9% 的成绩，展示了其在 STEM、人文学科和专业领域处理多步推理的能力。 这些性能提升源于一个核心设计选择：3.8 Flash 更加努力。在复杂任务上，它表现出更高的勤勉度——执行额外的推理步骤，并迭代调用工具。有时，模型可能会使用更多 token 以最大化性能，尤其是在更高的努力级别下。 对于计算效率是主要约束的应用，开发者可以利用较低的努力级别来最小化 token 开销，或继续依赖 Gemini 3.7 Flash，该模型仍完全支持效率优先的工作负载。 G

匹配字段正文

erted into an intermediate representation (e.g. code) before it can do anything. That intermediate representation limits what an interface can be, and how it responds to human and agent interaction. Every behavior has to be explicitly defined and implemented ahead of time, so software ships as a lossy compression of the space of possible interactions, frozen before any user arrives. The same translation process also sacrifices visual fidelity. Once a design is reduced to a simplified representat

匹配字段正文

erprise documents. It converts complex, multimodal files into structured, machine-readable data that can power enterprise knowledge use cases–including document indexing, RAG, and agentic retrieval. More than text recognition, Parse detects and understands key visual elements - such as tables and embedded images - and returns clean Markdown files for downstream processing and application. Use Parse to process your documents and images across nine major world languages. Parse is designed to prese

匹配字段正文

gine with autonomous surveys, helping teams turn vague feedback into named, fixable problems. With Unwrap, you can: Probe every open response with real-time AI follow-ups, so "the agent could've been better" becomes specific and actionable Auto-fire targeted surveys in response to feedback gathered on any channel Close the loop on customer feedback, turning churn risk into evangelists Stand up your own autonomous research program in minutes with a zero-cost, zero-commitment trial . The Rundown:

匹配字段正文

wen3.8-27B Qwen3.7-Plus DeepSeek-V4-Flash-0731 Claude-Opus-4.6 (Max) Params 125B 27B 397B 284B -- Activated params 6B 27B 17B 13B -- N-gram embedding params 51B -- -- -- -- Coding Agentic coding DeepSWE 1.1 58.7 42.2 16.5 54.4 -- Agentic coding SWE-bench Pro 62.5 61.7 55.8 56.0 53.4 Multilingual software engineering SWE-bench Multilingual 81.0 73.8 75.8 -- 77.5 Repo-level code generation NL2Repo-Bench 48.1 42.3 41.1 54.2 47.6 Agent Long-horizon office work CoWorkBench 73.9 70.7 65.1 45.1 68.2 Pr

匹配字段正文

its own scale. All three are strong reasoners and instruction followers. The clearest capability split shows up in post-training. The 8B and 30B models additionally go through an agentic RL block that teaches them to operate as agents: calling tools, editing and running code, driving a terminal, and searching the web inside real environments. Every model supports native tool calling. Served through an OpenAI-compatible endpoint (for example, with vLLM), it emits tool calls in the OpenAI functio

匹配字段正文

eding the throughput offered by most proprietary cloud models (around 140 tok/s, depending on the dataset). Our main goal for LFM2.5-2.6B was to make it the first viable on-device agentic model. In agentic workloads, the model reasons before every tool call, and the user waits through it all. That is where speculation pays most. In Figure 2, we show the final impact on latency by testing function calling on the BFCL dataset [4]. Across various multi-tool scenarios, DSpark reduces the latency by

匹配字段正文

一个可继续训练、拓展并适配真实场景的开放底座。 在下表的评测中，模型在代码、复杂推理和长上下文方向表现突出。相比总参数量约为其 2–3 倍的 base model，模型的整体表现仍具优势。 这使 flash-base 适合用于以下方向的继续训练与探索： 面向软件研发和大型代码库的代码模型； 需要多步分析与复杂推理的专业模型； 面向长文档、长流程和长程任务的 Agent 基础模型； 金融、医疗、科研、工业等专业领域的持续预训练与后训练。 Benchmark 描述的是特定设置下的 Base Model 能力，并不等同于聊天体验、Agent 执行效果或生产系统性能。Tiny 和 Flash 在知识、数学等部分评测上也各有能力边界，我们希望社区以此作为选择训练底座的参考。 03 为什么要开放 Base Model？ 真实世界不只有一种任务，也不只有一种正确答案。 真实世界任务往往具有长期目标、专业知识、复杂环境、工具调用和责任边界。医疗、金融、科研、工业和个人助手需要不同的数据、反馈机制和安全约束。一个官方后训练模型可以提供通用能力，却无法替代每个领域基于真实任务进行的训练、对齐和验证。 Ba

匹配字段正文

as surpassed Moonshot AI’s Kimi K3 and on some it’s surpassed Claude Fable 5 or GPT-5.6-Sol. Here’s a more complete comparison: This puts the model more or less at the frontier of agentic coding benchmarks, with only 750B parameters – a third of Kimi K3! The Z.ai blog post is rather straightforward, and starts with a bold sentence: Scaling post-training is all we did for GLM-5.3. GLM-5.3 is the same base model as GLM-5.2 with substantially extended post-training. To risk a broad oversimplificati

匹配字段正文