Moonshot 新发的 Kimi-K3.1-Pro 有 3T 参数,写代码跑分能跟 OpenAI、Anthropic 旗舰掰手腕,还带了个测 CAD 的新基准,演示挺炫。
Latest
Moonshot 的新 Kimi K3.1 Pro 在 14 项基准里 6 项压过美国前沿模型,价格还只有 Opus 5.5 的一半,可以关注。
月之暗面刚发的 Kimi K3.1 Pro,在 Terminal-Bench 4.0 上跑赢了 Claude Opus 5.5,成本只要三分之一,开源可自部署。
Lambert 简评 Kimi k3.1 pro 跑分,说 Moonshot 这回把 RL 后训练做好了,想了解中美模型差距现状的可以看看。
Moonshot 放出 Kimi 3.1 Pro,开放权重模型里基准分最高,11 月初开源权重,定价细节值得对比着看。
DeepSeek 新版开源模型在网络防御榜单上拿了开放模型第一,总分只差 GPT-6 受限版 12 分,价格还便宜 4 倍。
Kimi K3.1 Pro 参数 2.8T、上下文 1M,官方说编程和 agent 能对标 Opus 5.5,11 月 4 日还开放权重,可以等实测。
Kimi 又出新模型 K3.1 Pro,2.8T 参数 MoE,能直接做 Unreal 和 Blender 项目,金融场景幻觉降了 65%,还有个快 2-3 倍的版本可以试。
The interesting signal is that Moonshot's frontier claim is being benchmarked not just on raw scores but on token efficiency, which determines real serving cost. Builders choosing between Kimi and Zhipu's GLM 5.3 get a direct head-to-head data point. Watch for official benchmark numbers and pricing to confirm the efficiency argument.
Consumer-GPU inference lowers the hardware barrier that keeps most world models out of hobbyist and lab budgets, and LoRA hooks make it a practical base for robotics simulation. Developers building interactive video, embodied AI or game prototyping can grab the weights from Hugging Face and fine-tune on their own data. Compared with closed world models behind APIs, Hy-Nomas gives full control over generation pipelines.
MiniMax 把 M3.1-Flash 开放权重了,编程基准 DeepSWE 拿了 73.7%,还比 M3 便宜 70%,跑智能体的可以试试。
月之暗面的 Kimi K3.1 Pro 已经能在网页和 App 上用到了,主打复杂任务,感兴趣可以去试试手感。
Distilling a large teacher model's embodied demonstrations into a 27B open-source model cuts task costs by an order of magnitude, making advanced robot reasoning affordable to run on local hardware. Robotics teams and embodied-AI researchers can fine-tune or deploy RoboRSI-27B-Preview without relying on the proprietary teacher, and Chinese-origin open weights give ASEAN labs an accessible alternative to US-only robotics stacks.
一位博主拿 1017 和 Opus 5.5、Astra、Sol 同题对比,结果前者把后者都比下去了,还猜是蒸馏的功劳,可以看看他的具体任务截图。
SiliconFlow 拿 DeepSeek V4.1 Flash 和 GPT-6 Astra 跑同一个 Three.js 3D 场景提示词,效果打平还便宜一半,提示词直接在评论区自取。
MiniMax 那个 428B 的 M3.1 Flash 出 GGUF 了,有 256GB 内存的机器可以下载 203.5GB 的 Q3_K_M 版本本地跑。
有人爆料 Kimi K3.1 快来了,主打 3D 渲染,据说能跟 Fable 5.5 掰手腕,做 3D 的可以盯一下。
有人发现好几个团队都选了 Hermes 微调加 DeepSeek 底座这条路,趋同到发推调侃,做开源微调的可以看看。
把 AI 的纠错故意当没听见,FUMO 的 Fusion Code 还有 90% 会再提醒,其他型号只有七成左右,测法挺有意思。
快手旗下的 Kling 更新到 4.0 了,一次能塞 15 个参考素材,单条生成最长 30 秒还能设 10 个关键帧,做视频的可以去试试。
就是之前在 OpenRouter 上神秘出没的 Space Bunny,428B 总参只激活 23B,权重已放到 HuggingFace 可下载。
有人把 V4.1-Flash 去审查后跑分,结果挺有意思:抽象代数掉分,大学数学反而涨了,可以看看具体数字。
一天的大事都在这:MiniMax开源了428B的M3.1 Flash,OpenAI的GPT-6 Luna API便宜10倍速,还有Kimi K31要来了。
Local hardware requirements are the practical headline: one DGX Spark, H200 or GB200 means teams can keep agentic coding inside their own infrastructure instead of paying API costs. Builders running open-source agent stacks should compare it against Chinese open models like DeepSeek on agentic coding benchmarks before choosing a default.
Production-grade serving of a MiniMax model through vLLM signals that the open-source inference stack can now carry commercial workloads with SLA guarantees. Teams evaluating cheaper open-weight serving options can look at Inferact's vLLM-based offering as a reference setup. Upstream contributions also mean the wider vLLM community inherits the reliability hardening that came out of this deployment.
DeepSeek 编程成绩在不同基准上对不上,TerminalBench 上换 agent 框架得分差 14 分多,测模型前可以看看这篇的坑。
The benchmark mix matters: it beats larger frontier models on diagnosis-style multiple-choice tasks while staying slightly behind on expert-level text reasoning, so it is a specialist rather than a general replacement. Developers and medical researchers can self-deploy and fine-tune it for literature research, clinical guideline retrieval, drug fact-checking and evidence curation apps. As an open-weight 124B MoE model, it lowers the cost barrier for healthcare AI teams in ASEAN and Chinese-speaking markets.
DeepSeek 和 MiniMax 的两个新模型都拿到了开源许可,想跟进国产开源动态的可以看看审批速度。
MiniMax又放了个开源模型M3.1-Flash,428B总参但只激活23B,跑起来比DeepSeek Flash的763B轻多了,自己部署的可以看看。
MiniMax 的 M3.1 Flash 权重开放了,不想自己管 GPU 的话,Together AI 上已经能直接用。
MiniMax 新出的 M3.1 Flash,主打代码库工程和多步自动化,vLLM 发布当天就能跑,还带 NVFP4 量化和投机解码,自部署的可以试试。
有人扒出 M3.1 Flash 疑似就是 M3 改了个名,只稀疏了几层注意力再加量化,关注模型版本套路的话可以看看。
Day-0 availability on a US inference platform signals Chinese labs continue to route distribution through Western providers to reach global developers. Teams building agents or long-context coding workflows can test a 1M-context model with vision on Together AI immediately, alongside alternatives like DeepSeek and Qwen hosted there.
MiniMax M3.1 Flash 发布当天就能用 SGLang 跑推理和 RL 训练,NVIDIA、AMD 卡都适配,自己部署的话可以关注下。
MiniMax 的 M3.1-Flash 发布当天就能用 NVIDIA NeMo 做微调了,要做 post-train 的开发者可以省去等适配的时间。
MiniMax 把 M3.1 Flash 权重放出来了,428B 只激活 23B,代理和编码基准分都贴出来了,想本地部署的可以去 Hugging Face 拉 weights。
V4.1 Flash-1017 在 ARC-AGI-2 上小胜 Gemini 3.1 Pro Preview,想看推理模型基准对比的可以看看这组数据。
The v2.0 changes make rankings harder to game: private test sets, anti-solution scanning and multi-vote LLM judging push mean scores down 1.7 points across 21 shared configurations. Teams choosing a coding agent should note that Opus 5.5's edge is concentrated on FrontierCode (56.0% versus 52.6% for DeepSeek V4.1 Flash), while 15 of 24 configurations cluster between 43% and 49%, so harness choice can matter as much as model choice.
Only the latest 2000 items are listed; for older ones see the daily archive