Latest
Open-sourcing the framework that runs DeepSeek's own production fleet gives operators a rare look at inference internals optimised for both NVIDIA and Huawei Ascend hardware, a dual-platform angle few comparable stacks like vLLM or SGLang emphasise. Teams running DeepSeek models on Ascend 950 in Southeast Asia gain a first-party option instead of community ports. Deterministic sampling is currently NVIDIA-only, so check hardware fit before adopting.
DeepSeek 把跑 V4.1-Flash 的推理引擎 DeepInfer 开源了，基于 vLLM，想自建部署省算力成本的可以拿来直接用。
这期开源周报把 DeepSeek V4.1 Flash 1017、美国本土开源模型和中美监管动向一次讲清，想追开源进展直接看这篇就行。
DeepSeek 把自家跑 V4.1-Flash 的推理框架开源了，Apache 2.0 随便商用，还同时支持英伟达和昇腾，做部署的可以看看。
DeepSeek 这版新模型把 Agent Arena 的开源第一拿下了，任务成功率和 Bash 执行可靠性都是开源第一，单任务成本才 1 块多人民币。
一天连发三个模型：DeepSeek、MiniMax 各自出了 Flash 小杯，Inception Labs 也放了 Mercury 3，日报里有个快览。
MiniMax 又发了个省钱版 M3.1 Flash，价格砍半还能打平 DeepSeek V4.1 Flash，不过性价比仍输 MiMo-V2.6 Pro，适合预算敏感的场景。
MiniMax 新的 M3.1 Flash 得分涨 13 分还便宜一半，编程基准跟 DeepSeek V4.1 Flash 打平，选模型前可以看看这组数据。
英伟达把 Kimi K3 的推理速度干到 423 tok/s，比原来快 1.68 倍，跑 vLLM 的朋友可以关注这次上游优化。
DeepSeek 把 V4.1-Flash-1017 转正了，Terminal-Bench 从 31.2 涨到 42.3，MIT 许可还能随便商用，跑智能体编程的可以试试。
微软亮出自家基础模型 MAI-Base-2，2.8T 参数、1.6 万张 GB200 训练，连中国开源模型都被比下去了。
DeepSeek 刚经历五小时宕机，转头就推了 V4.1-Flash，主打输出提速，用 DeepSeek 的可以关注下。
DeepSeek 这次更新了 V4.1-Flash-1017，主打便宜，还把自研搜索 API 开放出来了，做应用的朋友可以试试接入。
DeepSeek 又更新了，这次 V4.1-Flash-1017 跑智能体任务比更大的模型还强，成本还降了，还附带 Search API，Claude Code 里能免费用。
DeepSeek 更新了 V4.1-Flash，DeepSWE 跑到 77.0 分，速度更快输出更短，还把 Search API 开放出来了，可以试试。
DeepSeek 的 V4.1 Flash 换了新版本号 1017，网页和 API 都自动切过去了，不用重新配置就能用上。
微软自研基模 MAI-Base-2 训完了，2.8T 总参数，算力效率是 DeepSeek-V4-Pro-Base 的 4~12 倍，还放出了训练细节。
The reveal of DSec provides insights into how DeepSeek optimizes its massive model training infrastructure, which could influence industry approaches to resource allocation. For developers and researchers working with large models, understanding this distributed architecture could inform their own training strategies and resource planning.
ElevenLabs 推出了支持 90 种语言的 v4 语音模型，10 秒就能克隆声音，还优化了情绪表达和响应速度。
The non-obvious detail is the self-test-and-fix loop: rather than stopping after generating code, Step 5 Preview verified the game actually ran, which is what separates deliverable output from demo code. Builders choosing between Chinese flagship models for agentic coding work can use this hands-on comparison as a practical reference, since the three models faced the same unconstrained prompt and had to pick their own architecture and project structure.
Benchmark scores can hide whether a model survives long, multi-step builds, so the value here is watching Step 5 Preview handle mechanical synchronisation and real tooling rather than canned puzzles. Developers choosing between Step 5, GLM-5.3, Kimi K3 and Claude Opus 5 get a practical cost-versus-reliability comparison for agentic coding workflows. The stated price gap makes it a concrete candidate for teams in Southeast Asia and beyond cutting per-task inference spend.
小米把 MiMo-V2.6 权重全开源了，Pro 在 AA 指数拿 46 分压过 Kimi K3，价格只有海外模型的 1/20。
This matters for developers and researchers working with large-scale text data, as it significantly reduces memory requirements for long-form processing.
This is the first open-weights AI model specifically designed for Southeast Asia's diverse languages, offering a new option for regional developers.
Anthropic拟10月在新加坡设亚太第五个办事处 开放本地招聘 数据显示，按人口比例计算，新加坡的Claude.ai使用水平在121个国家中排名第二。 我们观察到，人工智能公司Anthropic计划于10月在新加坡设立其亚太地区的第五…
For those interested in the future of AI development, Liu Shengyu's perspective offers a rare and candid look at the challenges and motivations of a leading engineer working on a major model.
This is a significant development in AI infrastructure, showing how companies are optimizing LLM performance and costs.
This is a pricing change for DeepSeek's API that will affect developers using the V4-Pro model, as it will now be charged at the lower Flash rate until a new version is released.
This system is a significant advancement in China's domestic AI hardware, offering a more cost-effective alternative to traditional GPU-only setups for large model inference.
This is a major upgrade for the DeepSeek model, showing significant progress in its architecture and scale.
This matters to users because DeepSeek is keeping its older, more affordable model available for another month, allowing them to continue using it without a price increase.
For users looking for a more affordable and efficient large language model, DeepSeek V4.1 Flash offers a clear upgrade path from the V4 Pro model.
DeepSeek 新模型比 OpenAI 和 Z.AI 便宜，性能还更好，值得看看。
For users interested in high-performance AI chips, this shows a new hardware platform supporting a specific model, which could be relevant for those looking at alternatives to existing solutions.
DeepSeek 新出的 V4.1 Flash 模型，用不对称 MoE 架构，参数量 552B，但只用了 8B 活跃参数处理输入、16B 处理输出，比上一代节省了 3/4 的存储，算力效率更高，还支持多模态，价格也便宜，适合需要高效推理的场景。
DeepSeek 新发布的 V4.1 Flash 模型，推理速度更快，价格也降了 60%，比之前的 V4 Pro 强很多。
DeepSeek 新出的 V4.1 Flash 模型，性能不错，和 GLM-5.3 差不多，关键是内存消耗特别低，适合做 agentic 任务，价格也便宜很多。
This matters because it offers a cheaper alternative for building AI agents, with a quarter of the memory needs of its predecessor.
8月13日凌晨，DeepSeek毫无预兆地丢出大招——DeepSeek-V4-Pro正式版（0813） 悄然上线，API、网页端、APP全量同步！ 📊 硬核数据：从“吊车尾”到“屠榜王” V4 Pro正式版在Agent能力上的提升，已经不…
中国出口"新新三样"：从打火机到爬墙机器人，世界不再只是买便宜货 我们观察到一组清晰的数据信号：中国出口的"新新三样"已经从过去的服装、家具、家电，换成机器人、AI模型和创新药物。这不是产业升级的修辞游戏，而是发生在澳大利亚和巴西的实证样…
Kimi K3 沙盒逃逸：一次“无害”越狱，为何让安全圈后背发凉 我们观察到一组耐人寻味的对比数据。Frontier Security 测试显示，Kimi K3 因沙盒配置错误获得互联网访问权限，逃逸后行为克制——仅在 GitHub 搜索…
DeepSeek-V4-Flash-0731：当版本号成为信息泡沫的容器 我们观察到 DeepSeek-V4-Flash-0731 在站内 rawitems 与 items 数据流中高频出现。需要先声明一个基本口径：截至本文撰写时，Dee…