与「moe」相关的搜索结果

I am actually disgusted by the timing of this drop. And this will get a bit more technical. Let’s do a real technical teardown of this "1,385-sigma Blackwell weight fingerprint" report that OpenAI and Anthropic are breathlessly promoting today, because banking on nobody reading past the headline is how DC narratives get manufactured. ​First, look at the actual CUDA maths: when PyTorch initialises model weights (torch.nn.init) with a fixed PRNG seed, curand/Philox parallelises random number generation across the GPU’s Streaming Multiprocessors (SMs). Because Hopper has 132 SMs and Blackwell has 148 or 152 SMs, the exact same seed lays out initial weights in a different spatial pattern depending on the GPU’s SM count. In huge MoE models, certain low-gradient null-spaces or unrouted slices preserve faint directional traces of those step-0 weights even after training. ​So yes, brute-forcing the seed and SM grid can show that a model’s step_0 weight initialisation script ran on a 148/152-SM Blackwell GPU. Now read what the summary explicitly admits in the fine print: 1. Initialisation is not full training: The report literally states that these signatures show how the weights were created, and "support (but do not prove) the inference that Blackwell also participated in model training". Initialising a model checkpoint takes seconds on a single rented dev node before loading the safetensors onto a standard Hopper cluster. The "2,000 to 3,000 Blackwell GPUs" and "$500 million" cluster numbers do not come from the weight fingerprint and Depue admits in his own thread that he simply "estimated" cluster sizes from release schedules and model details. ​2. Zero geolocation and zero proof of smuggling: The final sentence of the executive summary explicitly admits that the findings "do not establish the physical location of the chips, how they were acquired, or whether export control regulations were violated". Depue himself rented a cloud GB200 instance just to run his verification test. Commercial labs that are not on the Entity List renting bare-metal Blackwell nodes in offshore clouds (Singapore, Japan, or the Middle East) is completely legal under BIS rules and leaves the exact same SM fingerprint. Now ask yourself why a former OpenAI researcher dropped this on November 12, the exact day Jensen Huang was pushed out, with OpenAI and Anthropic immediately stepping in to "independently verify" it. US closed-source labs are watching their paid API moats get eaten alive by free Chinese open-weights models. Conflating a step-0 PRNG initialisation seed with a "$500M smuggled Blackwell cluster" hands DC hawks the exact pretext OpenAI and Anthropic want to regulate global cloud compute and kneecap open-source AI.

Edge0：只用2.5GB内存，在电脑和手机本地运行35B MoE模型 Edge0-AI团队开源edge0-infer推理引擎，平时把模型文件存放在固态硬盘里，通过预先训练好的路由模块提前预测下一步要调用的模块，在计算当前内容的同时后台提前把所需数据从硬盘读进内存。 这个方法太巧妙了，因为35B的MoE模型实际激活只有3B，所以提前预测要激活的3B，把对应的模块放到内存即可。 16GB内存的Mac电脑、iPhone17 Pro Max手机、安卓手机以及普通Windows电脑都能流畅运行35B的MoE模型。运行时设备内存占用稳定在2.5GB到3.4GB左右，在苹果M5芯片上每秒能生成约25个token，在手机上每秒也能输出11到13个token。 GitHub：

本地无审查大模型的地下军火库今天彻底齐活了 从 8GB 入门卡到 24GB 满血工作站，Hugging Face 三大破甲顶流（Huihui、HauhauCS、DavidAU）把 Qwen 家族最凶残的 5 款无甲神作全放出来了。不仅切掉了烦人的道德说教锁，还完整保留了多模态视觉（Vision）与 MTP 多 Token 极速预测。 直接按你手头的显卡显存对号入座抄作业： 01 全能六边形战神：Huihui Qwen3.8 27B Abliterated 目前本地综合实力最强的破甲王牌之一，逻辑智商在线，原生保留视觉识图能力。 · 16GB 显存：直接冲 UD-IQ4_XS 或 IQ4（约占 14.4GB，黄金甜点档） · 24GB 显存：直接上 Q4_K_M 或 Q5 满血稳跑 · 极简启动：ollama run huihui_ai/Qwen3.8-abliterated仓库地址： 02 暴力狂暴破甲：HauhauCS Qwen3.8 27B Uncensored Aggressive 如果你嫌常规消融不够狠，这款走的是极其激进的深度破甲路线，几乎零拒答，同样带 MTP 加速与多模态视觉。 · 16GB 显存：选 IQ4_XS（约占 15.7GB，刚好卡满显存极限） · 24GB 显存：直接上 Q4 / Q5 仓库地址： 03 极速 MoE 闪电炮：Huihui Qwen3.6 35B-A3B Abliterated MTP 35B 的大身板，但底层是 MoE 混合专家架构，每次推理仅激活约 3B 参数，吐字速度快到飞起，支持视觉。 · 12–16GB 显存：跑 IQ3 / IQ4 配合专家卸载（建议系统内存 32GB+ 兜底） · 24GB 显存：直接全塞进显存跑 Q4，丝滑起飞 仓库地址： 04 激进破甲 + 极速吞吐：HauhauCS Qwen3.6 35B-A3B Aggressive 把 HauhauCS 最凶狠的激进去审查手法塞进了 35B-A3B 的 MoE 架构里，既要完全听话不拒答，又要极致生成速度的兄弟选它。 · 16GB 显存：IQ3 / IQ4 + 专家卸载 · 24GB 显存：Q4_K_M（约 21GB）完美吃满单卡上限 仓库地址： 05 8G 显存小钢炮：Qwen3.5 9B Defiant / Heretic 系列 DavidAU 操刀的 NEO-IMATRIX 极限调校版，搭载 Heretic 消融算法与 MTP，小体积榨出恐怖表现，轻量本首选。 · 8GB 显存：毫无压力轻松秒开 · 12–16GB 显存：显存极其宽裕，剩下的空间直接把上下文长度拉爆 仓库地址： 省流选型指南： · 追求极致智商与多模态全能：无脑冲 01 或 02（27B Dense） · 追求秒级吐字速度与高吞吐：选 03 或 04（35B MoE） · 8G 显卡或轻薄本长文本：直接锁死 05（9B Heretic） 本地跑大模型最爽的就是彻底摆脱云端审核与封号焦虑，搞红队测试、角色扮演和私有 Agent 的兄弟建议先把这份军火库马住备用。 你现在主力机的显卡显存是多少 G？平时更喜欢跑满血 Dense 模型还是飞快的 MoE 架构？ #AI# #大模型# #Qwen# #开源# #Ollama# #本地部署#

Wild rumor coming out of China this morning on DeepSeek V5: DeepSeek is reportedly already pretraining V5 as a ~2 Trillion parameter ultra-sparse MoE model. And instead of training on NVIDIA GPUs, they are reportedly running the V5 pretraining run on Huawei's new Ascend 950 / 950DT chips with a team of Huawei kernel engineers physically working inside DeepSeek's offices to keep the cluster humming. If DeepSeek pulls off a 2-trillion-parameter frontier run natively on Ascend 950 silicon, that changes the entire AI hardware story.

I LOVE DOING IRL STREAMS LIKE THIS for legal reasons moe and i are unharmed and had so much fun in Barcelona:) moe was holding on to the camera and one tiny strap on the seat LOL

特别感谢 @Ion_Mio_ 的 moejs：一个纯 Go 写的 JavaScript 运行时，专门为快速跑大量小插件沙盒而做。magpie 的中间件就跑在它上面，所以能留在网关进程里，每个流式事件只花 1 微秒左右。

magpie 的插件现在不只能接订阅，还能当网关中间件。 中间件在 magpie 网关里处理每个 Agent 发出的请求和收到的回复，不管背后是哪个 provider。它在进程内运行（moejs），每个流式事件大约 1 微秒。 第一批社区中间件大多照搬 New API 的渠道设置，JSON 也一样： · param-override：参数覆盖 · model-map：模型重定向 · system-prompt：系统提示词 · word-guard：敏感词过滤 · think-tags：去掉回复里的 … 在「插件 › 发现 › 网关中间件」里点一下就能装，选项直接在那一行里改。卡片上也标明了哪个是 Provider，哪个是 Middleware。 自己写一个：