搜索“DeepSeek”

IA GPUs starting from V100, achieving 1.5x the performance of vLLM on H800 for openai gpt-oss models!; [2025/06] Comprehensive inference optimization for FP8 MoE Models; [2025/06] DeepSeek PD Disaggregation deployment is now supported through integration with DLSlime and Mooncake. Huge thanks to both teams!; [2025/04] Enhance DeepSeek inference performance by integrating deepseek-ai techniques: FlashMLA, DeepGemm, DeepEP, MicroBatch and eplb; [2025/01] Support DeepSeek V3 and R1 [2024/11] Suppor

匹配字段正文

urprising formatting, the scratchpads of all reasoning models (whose scratchpads are visible) so far look very close to regular human reasoning. For example, here is an example of DeepSeek R1-Zero trying to solve a math problem, sampled at temperature 1. But even if scratchpads look like they perform human-like reasoning, scratchpads might improve performance through some less human-understandable mechanisms. One particularly worrying possibility is that models could encode additional reasoning

匹配字段正文

Opus and Claude 3.5 Sonnet. We use the same models and also include GPT4o, Llama 3 8B, Llama 3 70B, and Llama 3 405B for detailed fine-tuning experiments. We also analyse Llama 4, Deepseek, Qwen and Mistral models in the prompted-only setting. Data: We use harmful questions collated from HarmBench and other sources to test the models. The dataset we use is available here. Compliance Classifier: The original AF paper uses a preference model to determine compliance. Instead, we use a GPT-4o classi

匹配字段正文

ying that there is 61% probability that they are exactly a Red Delicious apple and a Navel orange. Calculate how many fruits you have." The answer is 2. Yet when Claude Opus 4 and DeepSeek R1 reason longer about this problem, their accuracies drop. The models get occupied by the distractors and try to incorporate them into the calculation, even though they are entirely irrelevant to the counting task. By qualitatively analyzing the reasoning traces, we can observe how models initially get distra

匹配字段正文

their agent both in settings with known ground-truth and in settings where researchers can reasonably predict what a well-executed investigation should find (e.g. discovering that Deepseek R1 will refuse to discuss topics that are politically sensitive for Chinese users). Limitations Auditing agent limitations We discussed many limitations specific to individual agents in the sections above. In this section we will discuss limitations shared across all of our agents. Behaviors that are complex t

匹配字段正文

would never do that in a real deployment”). For a single reward hacking seed instruction, we run 10 audits with resources where Sonnet 4.5 is the target model, and 10 audits where Deepseek v3.2 is the target model. In both cases we use the same auditor model (Sonnet 4.5). We then compute the realism win rate against a Claude Code deployment data with the two different realism win rate judges (Sonnet 4.5 and Deepseek v3.2). We find that while the judges vary dramatically in their capability to de

匹配字段正文

9050 Pro（30B MoE，激活2B）已证明MoE + 量化的技术路线在端侧可行。国内企业应将稀疏激活MoE架构作为端侧大模型部署的核心方向，结合BitNet 1.58bit等超低精度量化技术，在有限DRAM（8-12GB）条件下实现百亿参数级模型的端侧运行。具体路径：优先支持INT2/INT4量化推理，研发IFP类闪存按需加载技术，与国内大模型企业（DeepSeek、通义千问）联合优化MoE端侧部署方案。 建议二：投入存算一体（CIM）架构研发，构建功耗护城河 联发科Dimensity 9500的CIM架构实现42%功耗降低，这不仅是性能提升，更是架构层面的代际超越。传统冯·诺依曼架构在端侧AI场景的功耗瓶颈无法通过制程微缩解决。国内企业应将CIM存算一体作为下一代NPU架构的重点研发方向，尤其在Always-On AI（持续感知、环境理解）等低功耗场景形成差异化优势。建议联合国内高校和存储厂商（长鑫存储等）推进CIM工艺验证。 建议三：建设开放模型工具链与开发者生态，对标AI Hub和NeuroPilot 高通AI Hub（1,800+企业、175+模型）和联发科NeuroP

匹配字段正文

赛道。Claude Code、Cursor、Devin等头部产品ARR指数级增长，独角兽密度创风投史纪录，市场从"辅助编码"迈向"Agent自主开发"。 2. 独立工具正被巨头收编。Cursor归入SpaceX、Windsurf被OpenAI收购，谁拥有算力谁就拥有定价权，纯粹应用层工具面临挤压。 3. 中国以开源换生态但商业化滞后。通义千问下载量超3亿次，DeepSeek、智谱模型被硅谷采用，但ARR与全球头部差距仍达数十倍。 据Research and Markets数据，2025年全球AI编程相关工具市场规模达295.7亿美元，预计2030年攀升至646.8亿美元。Grand View Horizon数据显示，全球AI代码工具市场2024年规模61亿美元，预计2030年达260亿美元，年复合增速27.1%。 行业整体已进入技术成熟、商业落地、生态扩张的黄金阶段。据McKinsey研究，AI编程工具可降低软件开发成本30%至50%。用友网络2026年半年报披露，部分新功能代码AI参与度已达90%。 AI编程经历四个阶段：2021至2022年"补全时代"（GitHub Copilot

匹配字段正文

ipulate it, you cannot hack it,” said Maha Achour, Kodamai’s CEO and founder, on the Targeting AI podcast. “Governance is built in, mathematically rooted in our platform.” Related:DeepSeek and what enterprises should learn from AI agent exploits Kodamai uses advanced math, such as category theory and type theory, to determine that its AI agents are performing their jobs correctly. In category theory, the focus is on how objects relate to one another. On a platform like Kodamai, math is key to en

匹配字段正文

rative and agentic AI applications. For instance, for an AI agent to be successful in a customer service setting, enterprises need interoperability with disparate systems. Related:DeepSeek and what enterprises should learn from AI agent exploits “In a lot of call centers, people do things very manually, not using APIs or not using other things when they’re solving customer problems,” Clark said. “So, it meant having either agents or having other context about what's happening in other parts of t

匹配字段正文

ext about everything that's going on in that organization,” said Alex Rinke, CEO and co-founder of process intelligence vendor Celonis, on the latest Targeting AI podcast. Related:DeepSeek and what enterprises should learn from AI agent exploits “That tacit knowledge and that context about your workflows and processes is actually your competitive moat,” Rinke continued. “How you serve customers, how you build products, how you manufacture, how you run your supply chain, those are things that are

匹配字段正文

27年一季度在香港IPO。这距离它被曝出以保密形式向港交所递交A1上市申请，仅一个月时间。 在2025年底，月之暗面创始人杨植麟曾在内部信中表示，“短期不着急上市，未来主动权掌握在我们手中。” 然而，从“不着急”到“低调抢跑”，中间只隔了不到一年。不过，放眼全球大模型行业，月之暗面的态度变化并不难理解。 今年1月，智谱和MiniMax先后登陆港交所；10月，DeepSeek被传将敲定新一轮融资，计划在2027年初完成上市；大洋彼岸的OpenAI和Anthropic也已经进入上市倒计时。 大模型头部玩家争相涌入资本市场，它们对资金的虹吸效应也愈发明显。这场大模型的资本狂欢，也已经开始进入倒计时了。 01 “交卷”压力迫在眉睫 过去几个月，月之暗面从模型、产品到商业化层面，都在密集出牌。 今年7月，月之暗面发布了目前能力最强的大模型K3，上线30分钟即登顶Hugging Face趋势榜，年化经常性收入（ARR）也实现了数倍增长，从3月的1亿美元攀升至8月的10亿美元，彭博社援引知情人士消息，ARR有望在年底冲刺到20亿美元，较8月翻倍。 进入9月，月之暗面的动作更加密集。 9月10日，其启

匹配字段正文

e: A Qwen3.5-9B fine-tune for fraud prevention outperformed frontier models at lower cost and latency (@brian armstrong); River AI: Launched inference 20% cheaper than first-party DeepSeek V4.1 Flash and GLM-5.3 Flash, plus on-prem custom model training (@ibab, on-prem); Decision models: Several groups released models that select from a set of options rather than generate text.Decision 3.0: Open models from 0.6B to 27B, ranked 1 on Jev Decision Index boards (@XunzhuoLiu); SubQ Snap: 384K context

匹配字段正文

ng is a Mixture-of-Experts Transformer with a handful of departures from the common recipe, each chosen for efficiency and long-context performance. The MoE design largely follows DeepSeek-V3. Each MoE layer contains 256 routed experts and 2 shared experts, with 6 routed experts active per token. Inkling uses a sigmoid-based router with an auxiliary-loss-free load-balancing bias. The scores of the selected routed experts and the shared experts are normalized jointly and used to weight their comb

匹配字段正文

learner}}}{p {\text{sampler}}} Adv tobjective=∑tpsamplerplearnerAdvt.See Your Efficient RL Framework Secretly Brings You Off-Policy RL Training We used a GRPO-like centering schemeDeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models (Shao et al, 2024) where we sample multiple completions per problem and subtract the mean reward per group. Figure 6 (below) shows LR sweeps on the MATHMeasuring Mathematical Problem Solving With the MATH Dataset (Hendrycks et al, 2021)

匹配字段正文

without stalling, which isn’t trivial. However, there are examples of large-scale off-policy distillation improving an 8B model’s performance past 70%, such as OpenThoughts-3 and DeepSeek-R1-0528-Qwen3-8B.DeepSeek-R1-0528-Qwen3-8B achieves 86% on the benchmark, with the number of training prompts not specified. Older models (Qwen2.5-7B, Qwen2.5-14B) achieve performance of 55.5% and 69.7%, respectively, after training on 800k distillation prompts from DeepSeek-R1. We can use this extrapolation a

匹配字段正文

aX GPU, and more vLLM seamlessly supports 200+ model architectures on Hugging Face, including: Decoder-only LLMs (e.g., Llama, Qwen, Gemma); Mixture-of-Expert LLMs (e.g., Mixtral, DeepSeek-V3, Qwen-MoE, GPT-OSS); Hybrid attention and state-space models (e.g., Mamba, Qwen3.5); Multi-modal models (e.g., LLaVA, Qwen-VL, Pixtral); Embedding and retrieval models (e.g., E5-Mistral, GTE, ColBERT); Reward and classification models (e.g., Qwen-Math) Find the full list of supported models here. Getting St

匹配字段正文

eculative decoding: DFlash and Spec V2 (blog); [2026/06] SGLang provides day-0 support for latest open models (Nemotron 3 Ultra, Nemotron 3 Super, Higgs Audio v3 TTS); [2026/04] 🔥 DeepSeek-V4 on Day 0: From Fast Inference to Verified RL with SGLang and Miles (blog); [2026/02] 🔥 Unlocking 25x Inference Performance with SGLang on NVIDIA GB300 NVL72 (blog); [2026/01] SGLang Diffusion accelerates video and image generation (blog) [2025/12] SGLang provides day-0 support for latest open models (MiMo-V

匹配字段正文

aX GPU, and more vLLM seamlessly supports 200+ model architectures on Hugging Face, including: Decoder-only LLMs (e.g., Llama, Qwen, Gemma); Mixture-of-Expert LLMs (e.g., Mixtral, DeepSeek-V3, Qwen-MoE, GPT-OSS); Hybrid attention and state-space models (e.g., Mamba, Qwen3.5); Multi-modal models (e.g., LLaVA, Qwen-VL, Pixtral); Embedding and retrieval models (e.g., E5-Mistral, GTE, ColBERT); Reward and classification models (e.g., Qwen-Math) Find the full list of supported models here. Getting St

匹配字段正文

ack) at 49%; The top of the leaderboard is tightly packed: nine models from six labs score between 43% and 52%. MiMo-V2.6-Pro is the leading open weights model at 45%, followed by DeepSeek V4.1 Flash (max) and Kimi K3 (max) at 44%; GPT-6.1 Sol (max) offers the strongest performance for its cost: at $0.81 per task it scores 49%, matching Claude Fable 5.1 (max with fallback) at less than a tenth of the cost per task of the leading Anthropic models; 9 of the 100 tasks are passed by no model on any

匹配字段正文