DeepSeek 把自家线上跑 V4.1 Flash 用的推理框架开源了,还兼容昇腾,做私有化部署的可以看看
Latest
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。
GEPA 在 MATH-500 上能从 0.828 拉到 0.904,但在 HotpotQA 上完全没用,SkillOpt 还会翻车——如果你在本地跑推理模型做提示词优化,这篇能帮你避坑。
Memory write-injection looks like a cheap way to add knowledge to a deployed model, but these results show the added facts cannot be chained into multi-hop reasoning by the model itself — orchestration must live in the application layer. Builders wiring RAG-style memory into agents should chain written atoms explicitly in prompts or system logic, and benchmark vendors should adopt coverage, synergy and distortion reporting. The tested beds include Chinese frontier models Qwen3 and DeepSeek, making the findings broadly relevant across model families.
拿四个 LLM 模拟欧洲四国调查,看谁能复现真实的年龄差距,结果三个模型翻车,适合做问卷模拟的人看看。
一篇研究智能体自我改进的论文,Qwen3.5-9B 用它这套框架跑赢了 DeepSeek-V4-Pro 和 GLM-5.3,做 agent 的可以看看思路。
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。
A 0.5B model beating much larger commercial systems on pronunciation suggests small, MIT-licensed TTS can now be self-hosted for production voice applications at minimal compute cost. Developers building voice agents or dubbing pipelines in the region can download both variants from Hugging Face and run them locally, avoiding per-character API fees from providers like ElevenLabs. The main trade-off is that closed models still lead overall, with Ming-3.0-tts only #13 on Controlled Voice.
有人实测 deepseek-v4.1-flash-1017,说 Agentic Coding 从 GROK-4.6 水平直接追上国产旗舰,还发现了行为变保守、预算控制的问题。
做 Agent 长任务的朋友看看这篇:把设计意图存图里、用确定性控制器核对交付物,19 个领域、228 个任务的基准测下来,同一模型约束满足率高出基线 14-25%。
Skills 之间怎么传递输出一直靠模型自己猜,这篇把依赖关系做成 DAG 直接可执行,做 Agent 编排的可以看看实现思路。
The same serving stack now optimises both NVIDIA and AMD paths, giving operators a concrete reason to benchmark MI355X deployments against GB300 rather than defaulting to CUDA. Teams running agentic inference traces can enable the mode through SGLang and expect P90 latency gains without changing models. The plugin system also means third-party kernels — starting with GLM-5.3, K3, and MiniMax M3 — will arrive without waiting for a core release.
医疗对话模型排行榜上线了,GPT-6 Astra 和 DeepSeek V4.1 Flash 的分数、成本对比都有,选医疗模型前可以看看。
DeepSeek 推了 V4.1 Flash 1017 新权重,Databricks 上能直接调了,写代码、跑智能体和看图任务都比之前强,用 Databricks 的可以试试。
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。
DeepSeek 这版 Flash 模型在 Agent Arena 拿下开源第一,跑任务成本只有 GPT-6.1 Sol 的四分之一左右,写网页还追平了闭源大模型,做智能体的可以看看。
DeepSeek 更新了 V4.1 Flash,智能指数涨到 45 分,追平 GLM-5.3 成开源第二,幻觉还少了好多,价格没变。
Fish Audio 的 Drama 3 冲上开源 TTS 第一名,发音还赢了 ElevenLabs v4,创始人还讲了怎么把发布帖做到 1300 万浏览。
Efficiency is the headline: the model used ~189M output tokens versus 324M for GLM-5.3-Flash, so low cost comes from token frugality as much as pricing. Its 93% hallucination rate on AA-Omniscience means teams should add verification layers for factual workloads. An open-weights release on 23 October 2026 would make it directly comparable with DeepSeek's open models for ASEAN startups building on local hosting.
Cache hit rate alone is a poor proxy for cost: models with nearly identical hit rates showed a more than seven-fold gap in what users actually paid versus list price, making DeepSeek's pricing structure unusually favourable for long agent tasks. Teams deploying office agents in China can use the per-model effective-cost figures to pick cheaper backends, with LobsterAI reachable via desktop, WeChat and Feishu. The report is self-reported product data, so treat the absolute numbers as one product's view rather than market-wide evidence.
DeepSeek 新版 V4.1-Flash-1017 上 SiliconFlow 了,DeepSWE 编码基准跑到 77%,跑智能体任务更省 token 也更便宜,可以试试。
Recraft 晒了一张 V4 Styles 生成的设计图,提示词直接放在评论区,做图像风格的可以去抄作业试试。
有人注意到 DeepSeek 和 MiniMax 的新模型参数反而变小了,算力吃紧可能正在改变训练策略,趋势挺有意思。
MiniMax 放出 M3.1 Flash 官方跑分,6 项基准里 4 项输给 DeepSeek V4.1 Flash,但高峰期价格便宜一半,选谁看场景。
Self-feedback reinforcement learning that targets whole-project output, not single-file answers, signals a shift in coding benchmarks toward multi-file, runnable deliverables. Developers evaluating open-weight models for end-to-end scaffolding can compare V4.1-Flash-1017 against Claude and GPT coding models on game and 3D tasks before committing to a stack.
有人发现 DeepSeek 这个代号奇怪的模型在 Terminal-Bench 4.0 上拿了 42.3,还比 GLM-5.3 强,等 Artificial Analysis 复测。
中国两部委出了个 AI 出口管制草案,2T 参数以上不让随便开源,Kimi K3 这类大模型直接中招,做开源的得留心了。
17万粉的SEO老哥发DeepSeek V4课程,俩小时才854曝光1个赞,做内容的朋友看了心里会平衡不少。
DeepSeek 新的 V4.1 Flash 1017 在 Together AI 能直接用了,专攻从零写完整 Web 应用和游戏,价格没涨,写代码的可以试试。
DeepSeek 新版 V4.1-Flash-1017 的第三方点评来了:幻觉控制最好,但知识和 LCR 倒退,还被 Sol 压着打。
网易 LobsterAI 拿真实用户数据写的报告:微信是最大入口,七成人只用一个模型,跟 X 上的玩法差别挺大,做 Agent 产品的可以看看。
有人把 DeepSeek V4.1F 和 Claude Opus 5.5 拉出来做前端设计对比,结论居然是 Opus 越改越差,有意思
rauchg 用 DeepSeek 新的 v4.1 flash fast 模型给小游戏加了音效、碰撞检测和跟随镜头,全程才花了 30 美分。
有人拿 Opus 5.5 写设计稿、DSv4.1 Flash 写代码,结果代码一堆没用的模块,Opus 也修不动,看看这个翻车组合。
DeepSeek 新版编程模型把 Kimi K3 挤下开源第一,成本只有十分之一,价格没涨,做 Web 开发的可以直接去 API 试试。
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。
有人质疑 V4.1 Pro 可能跟 Flash 差距不大,理由是 RL 预算和蒸馏数据卡了上限,观点挺有参考性。