搜索“DeepSeek”

中的任何一款。 在发布时可供使用的 API 提供商 Together AI、Fireworks 和 Baseten 处，每百万输入 token 收费 1.25 美元，每百万输出 token 收费 4.25 美元。尽管 token 单价相对较高，但得益于较低的消耗量，AA 计算出每个基准任务的平均成本为 0.84 美元。GLM-5.3 为 2.01 美元，而 DeepSeek V4.1 Flash 仅为 0.27 美元。 不同 AI 模型在性能与每个基准任务平均成本方面的比较。模型位置越靠左上，在性能与成本方面的表现越好。 （图：Artificial Analysis） 希望自行运行 Muse Spark Open 的用户，可从 Hugging Face 下载模型权重。模型权重提供 16 位版本（约 1.6 TByte）和部分量化为四位的版本（约 530 GByte）。 Muse Model License 原则上也允许商业使用，但要求标注“Built with Muse Spark Open”并遵守一项使用政策。通过 API 将 Muse Spark Open 作为服务提供的企业，若在

匹配字段正文

键值头，并对查询和键使用 RMSNorm。一个从层输入计算出的学习 sigmoid 门在输出投影之前缩放每个头的每个通道。 专家混合 前 3 层使用稠密 SwiGLU MLP。其余 49 层将其替换为包含 256 个专家的专家混合块，每个专家是一个中间大小为 2,560 的 SwiGLU MLP，约 39M 参数。专家占模型参数的 98.4%。 路由器遵循 DeepSeek-V3 的分组限制方案。它用 sigmoid 为所有 256 个专家打分，将它们分成 8 组、每组 32 个，保留最好的 4 组（按每组前两个分数之和排名），然后在其中选出前 8 个专家。Beam 使用与 DeepSeek-V3 相同的组大小，但没有共享专家。专家输出用归一化的路由器权重组合，并缩放 2.5 倍。 每个 token 只运行 8 个专家，因此模型用约 23B 参数计算每个 token。所有 256 个专家都保留在内存中，因此部署需要 501B 模型的内存，但以 23B 模型的速度运行。BF16 checkpoint 以 safetensors 格式约为 1 TB。NVFP4 checkpoint 为 N

匹配字段正文

超过 K2 Horizon（4.6%）和 Qwen3.8 27B（5.4%）两倍多。其最弱的结果是 Terminal-Bench 4.0（3.0%）和 AA-Briefcase（976 Elo） 在 AA-AgentPerf 上，NVIDIA Vera Rubin NVL72 每兆瓦的总吞吐量最高可达 GB300 NVL72 的约 5.7 倍，首批结果涵盖 DeepSeek V4 Pro、DeepSeek V4.1 Flash、GLM-5.2 和 Kimi K3。 2026年10月15日 NVIDIA RTX Spark N1X 笔记本电脑的首批 AA-AgentPerf-Local 结果 一台 NVIDIA RTX Spark N1X 笔记本电脑在 18.7 分钟内完成了我们在 Qwen3.8 27B 上的本地智能体工作负载。 2026年10月15日 NVIDIA 发布 Nemotron 3.5 Super VL 和 Nemotron 3.5 Bolt 74B A9B NVIDIA 的 Nemotron 3.5 Super VL 现已成为领先的美国开源权重模型，在 Artificia

匹配字段正文

it costs $0.84 per Intelligence Index task. That is less than Muse Spark 1.2 ($0.97) at the same per-token prices and less than half of GLM-5.3 (max, $2.01), but about three times DeepSeek V4.1 Flash (max, $0.27) ➤ A relative weakness in agentic work compared to Chinese open-weights leaders. Among the nine highest-scoring open-weights models, Muse Spark Open ranks last on the four agentic evaluations in the index but fourth on the other six combined, and it leads all open-weights models on AA-Om

匹配字段正文

was clearly on speed. Nvidia reports fourfold faster completion of 10,000 SWE-bench Pro tasks on four B200 GPUs, though its chart shows lower accuracy than Qwen3.8-Flash-Next and DeepSeek V4 Flash. Hybrid Mamba architecture, faster decoding As for how Nvidia built the model, it’s worth noting that the Nemotron coalition is sticking to the 3.5 family’s hybrid architecture, which combines Mamba-2, attention, and mixture-of-experts layers — but with the addition of a vision encoder. Nvidia also co

匹配字段正文

有了更深的推理路径，下一步是控制上下文带来的注意力计算开销。 当上下文长度达到100万token时，全注意力层会成为训练和推理共同的效率瓶颈。 阶跃星辰的解法是在全注意力路径上采用一种名为Sparse GQA的稀疏注意力机制，即在标准的分组查询注意力（GQA）中引入稀疏索引，模型不必对全部历史信息做注意力计算，而是先筛选、再计算。 这一思路并非阶跃首创，此前DeepSeek的DSA等工作已做过类似探索。 阶跃星辰称，其实际部署中需要跨越三道坎：筛选器本身的开销、筛选后碎片化计算的GPU利用率、以及筛选精度与数据读取量之间的权衡。 相比MQA（Multi-Query Attention，多查询注意力）和 MLA（Multi-Head Latent Attention，多头潜在注意力）结构，GQA其实会放大这三重矛盾。团队之所以选择GQA，是因为该结构在实验中的长上下文精度更好，且更利于多GPU并行推理。 稀疏GQA示意图1 但这三重矛盾也不能忽视，针对此问题，阶跃给出了Sparse GQA具体方案： 将相邻的历史位置压缩为“块” 来降低筛选器成本，块大小设为8，筛选与选择环节的成本因此下

匹配字段正文

为 Step 5 Preview 部署提供高效算力支撑。 沐曦股份表示，曦云 C 系列 GPU 已率先完成 Day 0 深度适配，以全栈自研算力底座，为国产 Agent 生态再添核心支撑，持续领跑国产大模型极速适配赛道。 据介绍，自 2025 年 12 月以来，沐曦股份已完成 43 个主流旗舰模型的 Day 0 适配。范围涵盖智谱、阿里千问、MiniMax、DeepSeek、阶跃星辰、腾讯混元等头部模型厂商，适配速度、数量、生态广度均位居行业前列。 硬件层面，曦云 C 系列 GPU 基于沐曦自研 XCORE 架构，搭载大容量高带宽显存与 MetaXLink 高速片间互联，具备混合精度算力与集群线性扩展能力，天然适配 MoE 模型专家调度、百万 Token 长上下文推理的算力与显存需求，单机多卡、大规模智算集群均可稳定承载。 软件层面，核心依托沐曦自研 MXMACA 全栈软件栈，覆盖底层驱动、MXCC 编译器、深度优化算子库、主流 AI 框架适配层，原生兼容主流 AI 开发生态，可将传统大模型适配周期从数周压缩至小时级，大幅降低模型迁移成本。 壁仞科技表示，本次 Day0 适配中，针对

匹配字段正文

output of the network Collections mentioned in this article 1 More Articles from our Blog Jun Kim, oMLX creator and maintainer, joins Hugging Face to support the MLX community +1 DeepSeek-V4: a million-token context that agents can actually use Models mentioned in this article 4 mlx-community/Muse-Spark-Open-816B-A42B-oQ3e mlx-community/Muse-Spark-Open-816B-A42B-oQ4e Datasets mentioned in this article 1 merve/meta-muse-spark-open-assets Papers mentioned in this article 1 Perception Encoder: The

匹配字段正文

ough still 28% fewer than Nemotron 3 Super (114k) NVIDIA Vera Rubin NVL72 serves up to 5.7x the total throughput per megawatt of GB300 NVL72 on AA-AgentPerf, with first results on DeepSeek V4 Pro, DeepSeek V4.1 Flash, GLM-5.2 and Kimi K3. October 15, 2026 First AA-AgentPerf-Local results for an NVIDIA RTX Spark N1X laptop An NVIDIA RTX Spark N1X laptop completes our local agent workload on Qwen3.8 27B in 18.7 minutes. October 15, 2026 Welcome to the era of embodied AI: Introducing AA-Robotics v0

匹配字段正文

ghtarrow$XX 设置下，Hy-MT2-1.8B、Hy-MT2-7B 和 Hy-MT2-30B-A3B 分别达到了 79.77、86.89 和 87.47，分别相当于 Gemini 3.1 Pro$^{\mathrm{T}}$ 的 89.9%、97.9% 和 98.6%。特别是在该设置下，Hy-MT2-7B 和 Hy-MT2-30B-A3B 表现优于 DeepSeek-V4-Pro、Kimi K2.6、Qwen3.5-397B-A17B 和 Gemma4-26B-A4B 等强劲的基线模型。与 Hy-MT1.5-7B 相比，Hy-MT2-7B 将 XCOMET-XXL 评分从 80.98 提升至 86.89，展示出在整体多语言翻译性能上的明显优势。 在 WMT25 上，Hy-MT2-7B 和 Hy-MT2-30B-A3B 同样表现强劲。Hy-MT2-7B 取得了 63.86/71.21/82.24 的成绩，而 Hy-MT2-30B-A3B 达到了 62.89/71.08/84.34。相比于评分为 61.59/68.85/75.91 的 Hy-MT1.5-7B，Hy-MT2-7B 在

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

gical capabilities in open source foundation models? To investigate this question, we compared gpt-oss-120b to other released open source models. At first, we primarily considered DeepSeek R1-0528. Partway through our process, the Qwen 3 Thinking and Kimi K2 models were released, and we added these to our comparison set. These evaluations confirmed that Qwen 3 Thinking and Kimi K2 have advanced to a level such that gpt-oss-120b does not significantly advance the state of the art on biosecurity-r

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 contex

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

d balancing to ensure stable, balanced expert utilization and healthy signal propagation through residuals. For expert utilization, we built on auxiliary-loss-free load balancing (DeepSeek-AI et al., 2024), introducing cosine decay of expert-bias updates to reduce routing perturbations later in training. Sequence-level balancing further encourages balanced expert utilization on data outside the pretraining distribution, preparing the model for the changing distribution of downstream RL. As a res

匹配字段正文

Tim反掰iPhone Duo被质疑：是为流量吗？ 8手搓动画《牛来》票房6300万 母子二人走向财富自由 9iPhone Duo开合动画火出圈 国产折叠屏已经开始抄作业了 10黄仁勋在台上接听特朗普电话 一部折叠屏手机也意外成为焦点 1 新研究发现母亲生育年龄影响后代的身体特征、寿命以及行为表现 阅读 (3529) 评论 (0) 2 苹果正式发布iPhone 18 Pro系列 强化系统级端侧AI 售价小幅调高 阅读 (1650) 评论 (0) 3 Switch模拟器迎来性能飞跃 开发工作却在巅峰时刻宣告终结 阅读 (1622) 评论 (0) 4时隔多年 摩拜299元押金可全额退了 5OpenClaw 2.0正式发布 一次简化安装的尝试最终重塑整个产品 6马斯克变性女儿床上大战机器人引热议 7Anthropic指控中国多家AI公司大规模蒸馏 阿里巴巴、月之暗面和DeepSeek被指涉数千万次交互 9倒数日·Days Matter Android版年度重大升级：透明相机+倒数日历+事件库 10中国科学家公布800座“飞翼”客机方案 或重新定义商业航空 created by ceallan

匹配字段正文

(with Python), we report the average of 16 runs (avg@16). For IMO-AnswerBench, we report the average of 8 runs (avg@8); Baselines: a. GPT-5, Claude-4.5-sonnet, Grok-4 results and DeepSeek-V3.2 results are quoted from the GPT-5 post, GPT-5 for Developers post, GPT-5 system card, claude-sonnet-4-5, grok-4, deepseek-v3.2, the public Terminal-Bench leaderboard (Terminus-2), the public Vals AI leaderboard and the artificialanalysis. Benchmarks for which no available public scores were re-tested unde

匹配字段正文

the most recent round of tool-related messages is retained; For BrowseComp, we report scores obtained with context management using the same discard-all strategy as Kimi K2.5 and DeepSeek-V3.2; For DeepSearchQA, no context management was applied to Kimi K2.6 tests, and tasks exceeding the supported context length were directly counted as failed. Scores for Claude Opus 4.6, GPT-5.4, and Gemini 3.1 Pro on DeepSearchQA are cited from the Claude Opus 4.7 System Card; For WideSearch, we report resul

匹配字段正文

r Verifier (KVV) to choose high-accuracy services. Details: https://www.kimi.com/blog/kimi-vendor-verifier Footnotes 1. General Testing Details We report results for Kimi K2.5 and DeepSeek-V3.2 with thinking mode enabled, Claude Opus 4.5 with extended thinking mode, GPT-5.2 with xhigh reasoning effort, and Gemini 3 Pro with a high thinking level. For vision benchmarks, we additionally report results for Qwen3-VL-235B-A22B-Thinking; Unless otherwise specified, all Kimi K2.5 experiments were condu

匹配字段正文