{"text": "大家好,我是大叔大。说个真事,我上周想微调一个本地模型,先在电脑上装环境,从UDR到Conda折腾到半夜还是报错,那会儿真想摔键盘。转头去云端,租显卡,跑了一周,账单贵得吓人,训练数据还不敢往上传,全程提心吊胆。所以今天这期,我就给你讲一个把这些问题一次性解决的桌面APP,Oncelot Desktop。简单说,它就是一个能让你在自己电脑上免费跑和训练大模型的工具。它是免费开源的原生桌面应用Mac,OS,Windows,Linux,三个平台都支持,训练提速两倍,显存省70%,精度还不打折扣,500多个模型都能跑能训,而且不用写一行代码。接下来咱们一步步看,它到底怎么用。先说痛点吧,你是不是也被这几道门槛卡过?第一道,装环境,想本地训练大模型,得先配UDA,装Conda,还要处理各种依赖,光这一步就能劝退一半人。第二道,显存不够,教程动辄说,要24GB显存,普通8G显卡的用户只能干瞪眼。第三道云,服务烧钱,按小时计费不说,训练数据还得上传到云端,敏感数据根本不敢放。第四道,微调方法分不清,lower QLOR全参微调参数怎么调,显存怎么算,完全是一头雾水。说实话,这几道门槛,哪个单拎出来都够折腾的。好在现在有一款桌面App,把这些门槛一次性全拆掉了。那这款桌面App,到底是什么来头?Uncelos Desktop是Uncelos团队推出的免费开源桌面板,基于Tory框架是真正的原生应用,不是网页套壳,启动快,占用低。它有几个挺实在的特点,一是原生桌面体验,不是网页套壳,二是数据不出电脑,训练推理,数据全在本地硬件上完成,三是入口很省心,对话,微调,部署。生成一个界面,全搞定,上手也特别简单。三步,第一步,下载安装,去官网下载对应系统的安装包。双击装上就能启动,第二步,选模型下载,顶部选模型,或者进模型中心,按自己设备选量化版本,自动下载,第三步直接对话,或者开始训练,零配置就能开聊。这几个功能里,最值得说的,就是它的无代码,微调,不用写一行代码,就能完成大模型微调。你可以直接把PDF,CSV,JSON,DOC,X这些文件拖进去,它会自动清洗,自动构建数据集,训练方式也很全。LORA,QLORA,全餐微调,预训练,强化学习,全都支持。训完之后,可以直接导出GGUF等格式,接到Alamalama,CPP这些推理引擎上就能用。算下来特别划算,训练提数两倍,显存省70%,精度还不损失。说白了,同样一张显卡,别人训一个模型的时间,你能训两个,还能多塞下三倍的大模型。这里提醒一句,训练的时候,把P大小,就是Batch Size,设成1,2,3,能明显降低,显存占用,别让它爆内存溢出。那无代码微调,对显存的要求高不高,答案很简单,门槛大幅跳水,靠QRA4bit量化,微调所需显存被压到了很低的水平。3B模型,只要3.5GB,7B模型,5GB,8B,只要6GB,14B,8.5GB,27B,22GB,70B,41GB。也就是说,训练7B模型,只要5GB显存,训练8B,只要6GB,一块主流终端显卡就能跑起来,笔记本读显也能玩。原理其实一句话,Lawr只训练模型权重里1%的保矩阵QRA,再把权重量化到4bit,进一步省75%的内存,这就是Encelos内存优化的核心思路。除了训练,它的对话能力也不含糊,本地模型也能打工具,工具调用准确率,提升了50%,还带自愈机制,调用失败,会自动检测,修复。从4,同时支持并行对话,几个任务同时聊,它还支持安全的代码执行,在沙箱里,运行Bass和Python,调用更稳,防止死循环,防止代码串话,权限这块也做得细,模型访问文件,联网执行命令都要经过你确认可以完全在安全沙箱里跑,也可以授权直接改文件。另外还有不限量的私人网络搜索,和一个深度研究模式,先规划,再查员,最后成文。引用都有出处,光自己对话还不够,它还能让你的AI agent直接用上本地模型,一条命令就能接入主流agent工具,模型随便换,数据不出门,用量不花钱。在终端里输入,onslots start,clawed,就能接入clawed code,输入onslots start codex,接入openai codex,还有hermes,openclaw,opencode,分别对应hermes agent,openclaw和opencode,这么做的好处,其实就三个。第一,模型随便换,本地quian,deepseek,glm,想用哪个换哪个,随时切换。第二,代码不外泄,项目代码只在本地处理,再也不怕敏感,代码传云端。第三,用量不花钱,本地推理免费随便用,跑满也不心疼。模型够不够新,也是大家关心的事,它的生态非常丰富,支持500多个模型的推理与训练,文本,视觉,语音,嵌入,图像视频扩散,全类型覆盖。团队和多家模型厂商深度合作,做到当天首发,当天支持,叫DayZero。简单说,新模型当天发布,当天就能跑,能训不用等适配,第一时间体验最新大模型,最新的旗舰模型都能跑。比如quian 3.8 27b,kimik3,deepseek v4,glm5.3,minimax h3,还有museglimmer,jammer4,本地已有的GGUF模型,也能自动识别。不用重新下载,也可以手动指定自定义目录。它还不只是文本模型,图像,视频,语音,通通能训,能生成。图像视频方面,支持Flux,Z-Image,LTX-1,这些模型还能训练自己的扩散。Laura,比如SDXL,FLUX 2.0,Q1,Image,生成速度也很快,用minimax h3,生成一段960x544,124帧,8部的视频从70多秒,压缩到13秒。语音方面,文字转语音,语音转文字,Whisper,Q1,3ASR,全部本地处理。一份工具,覆盖文本图像视频,语音四条生成线,训练生成,微调,转写都在同一个界面完成。训练完的模型,怎么部署,它给了你三种方式。第一种,局域网访问,直接开启LNN模式。局域网内任意设备都能访问,流量不出网,延迟低,断网也能用。第二种,公网HTPPS,用免费的Cloud Flare隧道,把本地模型安全暴露到公网手机上随时能访问。第三种,OpenAI兼容API,提供标准接口,任何支持这个协议的应用都能直接接入。你也可以用Uncelot Run命令,启动模型服务,指定模型和推理参数,比如温度,采样方式这些就能对外提供接口。而且还能连接云端Provider,OpenAI,Anthropic,Alama,LamaCPP,VLLM,本地云端同一个界面来回切换。最后,隐私和兼容性。很多人特别关心,这里都讲清楚。隐私方面,零摇测,不上传任何使用数据,完全离线,也能正常跑。只会检测本机GPU类型来适配运行,硬件要求也不高。Windows 10或11,Mac OS 12以上,Ubuntu 20.04以上都行。VDI,AMD,Intel,显卡加Mac都能训练,存CPU也能用。对话和数据处理,没问题。几个常见问题也顺带回答,已经下载的模型能不能用,能自动识别?必须高端显卡吗?不用CPU,Mac都能跑。支持OpenAI兼容API吗?支持推理,偶尔曼,把联网搜索和代码执行关掉,速度立刻回来。说到这,Unclosed Desk的OPPO的核心价值已经很清楚了,一句话,总结它把本地训练大模型从极客专属变成了人人都能上手。它的核心价值就三个词,本地化,免代码,全类型模型支持,这也是桌面端本地训练大模型的新选择。回顾一下,无代码微调,丢文件就能训。两倍提速70%,神险存,精度无损。一行命令,接入AI agent。局域网,军网,API,全能部署。零瑶测,全力线,数据不出门,适合谁用。想本地微调模型的开发者注重数据隐私的企业和个人,想省云费用的用户,还有刚接触大模型的新手都值得试试。感谢观看,我是大叔大,后续会持续更新,下期见。", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 1.16, "text": "大家好,我是大叔大。", "tokens": [50365, 15248, 11, 15914, 3582, 39936, 3582, 1543, 50423], "temperature": 0, "avg_logprob": -0.09886689927842882, "compression_ratio": 1.222972972972973, "no_speech_prob": 7.903010711129976e-12}, {"id": 1, "seek": 0, "start": 1.4000000000000001, "end": 9.16, "text": "说个真事,我上周想微调一个本地模型,先在电脑上装环境,从UDR到Conda折腾到半夜还是报错,那会儿真想摔键盘。", "tokens": [50435, 8090, 7549, 6303, 6973, 11, 1654, 5708, 37390, 7093, 39152, 8897, 225, 20182, 8802, 10928, 41908, 39823, 11, 10108, 3581, 42182, 27067, 239, 5708, 8083, 227, 8051, 107, 47885, 11, 35630, 52, 9301, 4511, 34, 12233, 7235, 246, 21184, 122, 4511, 30018, 30124, 45726, 49817, 29900, 11, 4184, 12949, 22933, 6303, 7093, 34783, 242, 23049, 106, 5419, 246, 1543, 50823], "temperature": 0, "avg_logprob": -0.09886689927842882, "compression_ratio": 1.222972972972973, "no_speech_prob": 7.903010711129976e-12}, {"id": 2, "seek": 0, "start": 9.56, "end": 15.3, "text": "转头去云端,租显卡,跑了一周,账单贵得吓人,训练数据还不敢往上传,全程提心吊胆。", "tokens": [50843, 17819, 105, 39862, 6734, 1369, 239, 11957, 107, 11, 8204, 253, 1431, 122, 32681, 11, 32585, 2289, 2257, 37390, 11, 18464, 99, 47446, 18464, 113, 5916, 2392, 241, 4035, 11, 7422, 255, 10115, 225, 33188, 26075, 106, 14852, 1960, 40205, 29510, 5708, 7384, 254, 11, 11319, 29649, 20949, 7945, 2392, 232, 19095, 228, 1543, 51130], "temperature": 0, "avg_logprob": -0.09886689927842882, "compression_ratio": 1.222972972972973, "no_speech_prob": 7.903010711129976e-12}, {"id": 3, "seek": 0, "start": 15.46, "end": 20.080000000000002, "text": "所以今天这期,我就给你讲一个把这些问题一次性解决的桌面APP,Oncelot Desktop。", "tokens": [51138, 7239, 12074, 5562, 16786, 11, 22020, 23197, 2166, 39255, 20182, 16075, 5562, 13824, 34069, 27505, 21686, 17278, 5676, 111, 1546, 162, 5475, 8833, 8749, 11, 11747, 4933, 310, 49044, 1543, 51369], "temperature": 0, "avg_logprob": -0.09886689927842882, "compression_ratio": 1.222972972972973, "no_speech_prob": 7.903010711129976e-12}, {"id": 4, "seek": 0, "start": 20.42, "end": 23.92, "text": "简单说,它就是一个能让你在自己电脑上免费跑和训练大模型的工具。", "tokens": [51386, 11249, 222, 47446, 8090, 11, 11284, 5620, 20182, 8225, 33650, 41410, 17645, 42182, 27067, 239, 5708, 2347, 235, 18464, 117, 32585, 12565, 7422, 255, 10115, 225, 3582, 41908, 39823, 1546, 23323, 39806, 1543, 51561], "temperature": 0, "avg_logprob": -0.09886689927842882, "compression_ratio": 1.222972972972973, "no_speech_prob": 7.903010711129976e-12}, {"id": 5, "seek": 2392, "start": 23.92, "end": 34.94, "text": "它是免费开源的原生桌面应用Mac,OS,Windows,Linux,三个平台都支持,训练提速两倍,显存省70%,精度还不打折扣,500多个模型都能跑能训,而且不用写一行代码。", "tokens": [50365, 11284, 1541, 2347, 235, 18464, 117, 18937, 47402, 1546, 19683, 8244, 162, 5475, 8833, 44297, 9254, 34355, 11, 4367, 11, 45813, 1509, 11, 43, 259, 2449, 11, 10960, 7549, 16716, 15433, 7182, 35488, 11, 7422, 255, 10115, 225, 20949, 31217, 36257, 35477, 11, 1431, 122, 39781, 2862, 223, 5867, 8923, 34910, 13127, 14852, 1960, 12467, 7235, 246, 3416, 96, 11, 7526, 6392, 7549, 41908, 39823, 7182, 8225, 32585, 8225, 7422, 255, 11, 22942, 24384, 5676, 247, 2257, 8082, 19105, 23230, 223, 1543, 50916], "temperature": 0, "avg_logprob": -0.06835325095685608, "compression_ratio": 1.1921296296296295, "no_speech_prob": 1.1706961788870984e-11}, {"id": 6, "seek": 2392, "start": 35.14, "end": 36.68, "text": "接下来咱们一步步看,它到底怎么用。", "tokens": [50926, 14468, 4438, 6912, 8975, 109, 9497, 2257, 31429, 31429, 4200, 11, 11284, 33883, 15282, 9254, 1543, 51003], "temperature": 0, "avg_logprob": -0.06835325095685608, "compression_ratio": 1.1921296296296295, "no_speech_prob": 1.1706961788870984e-11}, {"id": 7, "seek": 2392, "start": 36.78, "end": 39.18, "text": "先说痛点吧,你是不是也被这几道门槛卡过?", "tokens": [51008, 10108, 8090, 28737, 12579, 6062, 11, 2166, 23034, 6404, 23238, 5562, 6336, 254, 6025, 8259, 101, 24880, 249, 32681, 16866, 30, 51128], "temperature": 0, "avg_logprob": -0.06835325095685608, "compression_ratio": 1.1921296296296295, "no_speech_prob": 1.1706961788870984e-11}, {"id": 8, "seek": 2392, "start": 39.540000000000006, "end": 45.980000000000004, "text": "第一道,装环境,想本地训练大模型,得先配UDA,装Conda,还要处理各种依赖,光这一步就能劝退一半人。", "tokens": [51146, 18049, 6025, 11, 8083, 227, 8051, 107, 47885, 11, 7093, 8802, 10928, 7422, 255, 10115, 225, 3582, 41908, 39823, 11, 5916, 10108, 38846, 52, 7509, 11, 8083, 227, 34, 12233, 11, 14852, 4275, 1787, 226, 13876, 17516, 39810, 3254, 251, 5266, 244, 11, 20690, 5562, 2257, 31429, 3111, 8225, 5087, 251, 46361, 2257, 30018, 4035, 1543, 51468], "temperature": 0, "avg_logprob": -0.06835325095685608, "compression_ratio": 1.1921296296296295, "no_speech_prob": 1.1706961788870984e-11}, {"id": 9, "seek": 4598, "start": 45.98, "end": 51.5, "text": "第二道,显存不够,教程动辄说,要24GB显存,普通8G显卡的用户只能干瞪眼。", "tokens": [50365, 19693, 6025, 11, 1431, 122, 39781, 1960, 1787, 253, 11, 21936, 29649, 34961, 9830, 226, 8090, 11, 4275, 7911, 8769, 1431, 122, 39781, 11, 29993, 19550, 23, 38, 1431, 122, 32681, 1546, 9254, 1486, 115, 14003, 8225, 26111, 36733, 103, 25281, 1543, 50641], "temperature": 0, "avg_logprob": -0.08313096906037609, "compression_ratio": 1.2537313432835822, "no_speech_prob": 1.3008779817247351e-11}, {"id": 10, "seek": 4598, "start": 51.739999999999995, "end": 56.64, "text": "第三道云,服务烧钱,按小时计费不说,训练数据还得上传到云端,敏感数据根本不敢放。", "tokens": [50653, 35878, 6025, 1369, 239, 11, 27408, 5087, 94, 23661, 100, 39623, 11, 26613, 7322, 15729, 7422, 94, 18464, 117, 1960, 8090, 11, 7422, 255, 10115, 225, 33188, 26075, 106, 14852, 5916, 5708, 7384, 254, 4511, 1369, 239, 11957, 107, 11, 7017, 237, 9709, 33188, 26075, 106, 31337, 8802, 1960, 40205, 12744, 1543, 50898], "temperature": 0, "avg_logprob": -0.08313096906037609, "compression_ratio": 1.2537313432835822, "no_speech_prob": 1.3008779817247351e-11}, {"id": 11, "seek": 4598, "start": 56.919999999999995, "end": 62.66, "text": "第四道,微调方法分不清,lower QLOR全参微调参数怎么调,显存怎么算,完全是一头雾水。", "tokens": [50912, 7536, 19425, 6025, 11, 39152, 8897, 225, 9249, 11148, 6627, 1960, 21784, 11, 34598, 1249, 43, 2483, 11319, 2129, 224, 39152, 8897, 225, 2129, 224, 33188, 15282, 8897, 225, 11, 1431, 122, 39781, 15282, 19497, 11, 37100, 1541, 2257, 39862, 6306, 122, 15590, 1543, 51199], "temperature": 0, "avg_logprob": -0.08313096906037609, "compression_ratio": 1.2537313432835822, "no_speech_prob": 1.3008779817247351e-11}, {"id": 12, "seek": 4598, "start": 62.9, "end": 65.44, "text": "说实话,这几道门槛,哪个单拎出来都够折腾的。", "tokens": [51211, 8090, 24726, 21596, 11, 5562, 6336, 254, 6025, 8259, 101, 24880, 249, 11, 17028, 7549, 47446, 6852, 236, 44561, 7182, 1787, 253, 7235, 246, 21184, 122, 1546, 1543, 51338], "temperature": 0, "avg_logprob": -0.08313096906037609, "compression_ratio": 1.2537313432835822, "no_speech_prob": 1.3008779817247351e-11}, {"id": 13, "seek": 4598, "start": 65.44, "end": 68.66, "text": "好在现在有一款桌面App,把这些门槛一次性全拆掉了。", "tokens": [51338, 2131, 3581, 25040, 32241, 48798, 162, 5475, 8833, 9132, 11, 16075, 5562, 13824, 8259, 101, 24880, 249, 27505, 21686, 11319, 6852, 228, 29327, 2289, 1543, 51499], "temperature": 0, "avg_logprob": -0.08313096906037609, "compression_ratio": 1.2537313432835822, "no_speech_prob": 1.3008779817247351e-11}, {"id": 14, "seek": 4598, "start": 68.72, "end": 70.64, "text": "那这款桌面App,到底是什么来头?", "tokens": [51502, 4184, 5562, 48798, 162, 5475, 8833, 9132, 11, 33883, 1541, 10440, 6912, 39862, 30, 51598], "temperature": 0, "avg_logprob": -0.08313096906037609, "compression_ratio": 1.2537313432835822, "no_speech_prob": 1.3008779817247351e-11}, {"id": 15, "seek": 7064, "start": 70.64, "end": 78.4, "text": "Uncelos Desktop是Uncelos团队推出的免费开源桌面板,基于Tory框架是真正的原生应用,不是网页套壳,启动快,占用低。", "tokens": [50365, 12405, 4933, 329, 49044, 1541, 12405, 4933, 329, 3919, 95, 10034, 253, 33597, 7781, 1546, 2347, 235, 18464, 117, 18937, 47402, 162, 5475, 8833, 43664, 11, 26008, 37732, 51, 827, 21416, 228, 7360, 114, 1541, 6303, 15789, 1546, 19683, 8244, 44297, 9254, 11, 7296, 16469, 239, 10178, 113, 1881, 245, 12829, 111, 11, 2392, 107, 34961, 10251, 11, 5322, 254, 9254, 41377, 1543, 50753], "temperature": 0, "avg_logprob": -0.09248378565493186, "compression_ratio": 1.323607427055703, "no_speech_prob": 1.2470306905154604e-11}, {"id": 16, "seek": 7064, "start": 78.6, "end": 87.66, "text": "它有几个挺实在的特点,一是原生桌面体验,不是网页套壳,二是数据不出电脑,训练推理,数据全在本地硬件上完成,三是入口很省心,对话,微调,部署。", "tokens": [50763, 11284, 2412, 6336, 254, 7549, 41046, 24726, 3581, 1546, 17682, 12579, 11, 2257, 1541, 19683, 8244, 162, 5475, 8833, 29485, 49657, 234, 11, 7296, 16469, 239, 10178, 113, 1881, 245, 12829, 111, 11, 11217, 1541, 33188, 26075, 106, 1960, 7781, 42182, 27067, 239, 11, 7422, 255, 10115, 225, 33597, 13876, 11, 33188, 26075, 106, 11319, 3581, 8802, 10928, 38114, 105, 20485, 5708, 41509, 11, 10960, 1541, 14028, 18144, 4563, 2862, 223, 7945, 11, 8713, 21596, 11, 39152, 8897, 225, 11, 13470, 16469, 110, 1543, 51216], "temperature": 0, "avg_logprob": -0.09248378565493186, "compression_ratio": 1.323607427055703, "no_speech_prob": 1.2470306905154604e-11}, {"id": 17, "seek": 7064, "start": 87.8, "end": 90.04, "text": "生成一个界面,全搞定,上手也特别简单。", "tokens": [51223, 8244, 11336, 20182, 21185, 8833, 11, 11319, 41816, 12088, 11, 5708, 11389, 6404, 17682, 18453, 11249, 222, 47446, 1543, 51335], "temperature": 0, "avg_logprob": -0.09248378565493186, "compression_ratio": 1.323607427055703, "no_speech_prob": 1.2470306905154604e-11}, {"id": 18, "seek": 7064, "start": 90.32, "end": 93.14, "text": "三步,第一步,下载安装,去官网下载对应系统的安装包。", "tokens": [51349, 10960, 31429, 11, 18049, 31429, 11, 4438, 17819, 121, 16206, 8083, 227, 11, 6734, 31929, 16469, 239, 4438, 17819, 121, 8713, 44297, 25368, 10115, 253, 1546, 16206, 8083, 227, 23305, 1543, 51490], "temperature": 0, "avg_logprob": -0.09248378565493186, "compression_ratio": 1.323607427055703, "no_speech_prob": 1.2470306905154604e-11}, {"id": 19, "seek": 9314, "start": 93.14, "end": 102.06, "text": "双击装上就能启动,第二步,选模型下载,顶部选模型,或者进模型中心,按自己设备选量化版本,自动下载,第三步直接对话,或者开始训练,零配置就能开聊。", "tokens": [50365, 2129, 234, 6336, 119, 8083, 227, 5708, 3111, 8225, 2392, 107, 34961, 11, 19693, 31429, 11, 2215, 231, 41908, 39823, 4438, 17819, 121, 11, 10178, 114, 13470, 2215, 231, 41908, 39823, 11, 31148, 36700, 41908, 39823, 5975, 7945, 11, 26613, 17645, 7422, 122, 1787, 229, 2215, 231, 26748, 23756, 42096, 8802, 11, 9722, 34961, 4438, 17819, 121, 11, 35878, 31429, 43297, 8713, 21596, 11, 31148, 45213, 7422, 255, 10115, 225, 11, 6306, 114, 38846, 34719, 3111, 8225, 18937, 40096, 1543, 50811], "temperature": 0, "avg_logprob": -0.08251486979257901, "compression_ratio": 1.3655913978494623, "no_speech_prob": 1.3639043887347935e-11}, {"id": 20, "seek": 9314, "start": 102.38, "end": 107.54, "text": "这几个功能里,最值得说的,就是它的无代码,微调,不用写一行代码,就能完成大模型微调。", "tokens": [50827, 5562, 6336, 254, 7549, 32311, 8225, 15759, 11, 8661, 40242, 5916, 8090, 1546, 11, 5620, 45224, 31364, 19105, 23230, 223, 11, 39152, 8897, 225, 11, 24384, 5676, 247, 2257, 8082, 19105, 23230, 223, 11, 3111, 8225, 41509, 3582, 41908, 39823, 39152, 8897, 225, 1543, 51085], "temperature": 0, "avg_logprob": -0.08251486979257901, "compression_ratio": 1.3655913978494623, "no_speech_prob": 1.3639043887347935e-11}, {"id": 21, "seek": 9314, "start": 107.94, "end": 114.92, "text": "你可以直接把PDF,CSV,JSON,DOC,X这些文件拖进去,它会自动清洗,自动构建数据集,训练方式也很全。", "tokens": [51105, 42766, 43297, 16075, 17349, 37, 11, 26283, 53, 11, 41, 10388, 11, 26649, 34, 11, 55, 5562, 13824, 17174, 20485, 6852, 244, 36700, 6734, 11, 11284, 12949, 9722, 34961, 21784, 43252, 11, 9722, 34961, 7360, 226, 34157, 33188, 26075, 106, 26020, 11, 7422, 255, 10115, 225, 9249, 27584, 6404, 4563, 11319, 1543, 51454], "temperature": 0, "avg_logprob": -0.08251486979257901, "compression_ratio": 1.3655913978494623, "no_speech_prob": 1.3639043887347935e-11}, {"id": 22, "seek": 9314, "start": 115.22, "end": 119.12, "text": "LORA,QLORA,全餐微调,预训练,强化学习,全都支持。", "tokens": [51469, 43, 2483, 32, 11, 13695, 46, 3750, 11, 11319, 22217, 238, 39152, 8897, 225, 11, 12501, 226, 7422, 255, 10115, 225, 11, 5702, 118, 23756, 29618, 2930, 254, 11, 11319, 7182, 35488, 1543, 51664], "temperature": 0, "avg_logprob": -0.08251486979257901, "compression_ratio": 1.3655913978494623, "no_speech_prob": 1.3639043887347935e-11}, {"id": 23, "seek": 11912, "start": 119.12, "end": 124.80000000000001, "text": "训完之后,可以直接导出GGUF等格式,接到Alamalama,CPP这些推理引擎上就能用。", "tokens": [50365, 7422, 255, 14128, 9574, 13547, 11, 6723, 43297, 4510, 120, 7781, 27561, 52, 37, 10187, 30921, 27584, 11, 14468, 4511, 32, 4326, 304, 2404, 11, 34, 17755, 5562, 13824, 33597, 13876, 41301, 13167, 236, 5708, 3111, 8225, 9254, 1543, 50649], "temperature": 0, "avg_logprob": -0.0745292287236372, "compression_ratio": 1.2114014251781473, "no_speech_prob": 1.3300971435370457e-11}, {"id": 24, "seek": 11912, "start": 125.26, "end": 129.4, "text": "算下来特别划算,训练提数两倍,显存省70%,精度还不损失。", "tokens": [50672, 19497, 4438, 6912, 17682, 18453, 2437, 240, 19497, 11, 7422, 255, 10115, 225, 20949, 33188, 36257, 35477, 11, 1431, 122, 39781, 2862, 223, 5867, 8923, 34910, 13127, 14852, 1960, 26075, 253, 28564, 1543, 50879], "temperature": 0, "avg_logprob": -0.0745292287236372, "compression_ratio": 1.2114014251781473, "no_speech_prob": 1.3300971435370457e-11}, {"id": 25, "seek": 11912, "start": 129.58, "end": 134.74, "text": "说白了,同样一张显卡,别人训一个模型的时间,你能训两个,还能多塞下三倍的大模型。", "tokens": [50888, 8090, 13558, 2289, 11, 13089, 14496, 2257, 44059, 1431, 122, 32681, 11, 18453, 4035, 7422, 255, 20182, 41908, 39823, 1546, 44848, 11, 2166, 8225, 7422, 255, 36257, 7549, 11, 14852, 8225, 6392, 13331, 252, 4438, 10960, 35477, 1546, 3582, 41908, 39823, 1543, 51146], "temperature": 0, "avg_logprob": -0.0745292287236372, "compression_ratio": 1.2114014251781473, "no_speech_prob": 1.3300971435370457e-11}, {"id": 26, "seek": 11912, "start": 135.12, "end": 141.92000000000002, "text": "这里提醒一句,训练的时候,把P大小,就是Batch Size,设成1,2,3,能明显降低,显存占用,别让它爆内存溢出。", "tokens": [51165, 35102, 20949, 40796, 2257, 34592, 11, 7422, 255, 10115, 225, 49873, 11, 16075, 47, 3582, 7322, 11, 5620, 33, 852, 35818, 11, 7422, 122, 11336, 16, 11, 17, 11, 18, 11, 8225, 11100, 1431, 122, 47421, 41377, 11, 1431, 122, 39781, 5322, 254, 9254, 11, 18453, 33650, 11284, 45827, 34742, 39781, 16084, 95, 7781, 1543, 51505], "temperature": 0, "avg_logprob": -0.0745292287236372, "compression_ratio": 1.2114014251781473, "no_speech_prob": 1.3300971435370457e-11}, {"id": 27, "seek": 14192, "start": 141.92, "end": 149.11999999999998, "text": "那无代码微调,对显存的要求高不高,答案很简单,门槛大幅跳水,靠QRA4bit量化,微调所需显存被压到了很低的水平。", "tokens": [50365, 4184, 31364, 19105, 23230, 223, 39152, 8897, 225, 11, 8713, 1431, 122, 39781, 1546, 4275, 32718, 12979, 1960, 12979, 11, 28223, 28899, 4563, 11249, 222, 47446, 11, 8259, 101, 24880, 249, 3582, 3509, 227, 29366, 15590, 11, 5363, 254, 48, 3750, 19, 5260, 26748, 23756, 11, 39152, 8897, 225, 5966, 32535, 1431, 122, 39781, 23238, 5014, 233, 21381, 4563, 41377, 1546, 15590, 16716, 1543, 50725], "temperature": 0, "avg_logprob": -0.0632207468486153, "compression_ratio": 1.308139534883721, "no_speech_prob": 1.544039028578137e-11}, {"id": 28, "seek": 14192, "start": 149.32, "end": 156.07999999999998, "text": "3B模型,只要3.5GB,7B模型,5GB,8B,只要6GB,14B,8.5GB,27B,22GB,70B,41GB。", "tokens": [50735, 18, 33, 41908, 39823, 11, 14003, 4275, 18, 13, 20, 8769, 11, 22, 33, 41908, 39823, 11, 20, 8769, 11, 23, 33, 11, 14003, 4275, 21, 8769, 11, 7271, 33, 11, 23, 13, 20, 8769, 11, 10076, 33, 11, 7490, 8769, 11, 5867, 33, 11, 17344, 8769, 1543, 51073], "temperature": 0, "avg_logprob": -0.0632207468486153, "compression_ratio": 1.308139534883721, "no_speech_prob": 1.544039028578137e-11}, {"id": 29, "seek": 14192, "start": 156.23999999999998, "end": 162.04, "text": "也就是说,训练7B模型,只要5GB显存,训练8B,只要6GB,一块主流终端显卡就能跑起来,笔记本读显也能玩。", "tokens": [51081, 6404, 5620, 8090, 11, 7422, 255, 10115, 225, 22, 33, 41908, 39823, 11, 14003, 4275, 20, 8769, 1431, 122, 39781, 11, 7422, 255, 10115, 225, 23, 33, 11, 14003, 4275, 21, 8769, 11, 2257, 47734, 13557, 27854, 10115, 230, 11957, 107, 1431, 122, 32681, 3111, 8225, 32585, 9147, 6912, 11, 5437, 242, 34756, 8802, 5233, 119, 1431, 122, 6404, 8225, 19912, 1543, 51371], "temperature": 0, "avg_logprob": -0.0632207468486153, "compression_ratio": 1.308139534883721, "no_speech_prob": 1.544039028578137e-11}, {"id": 30, "seek": 16204, "start": 162.04, "end": 170.57999999999998, "text": "原理其实一句话,Lawr只训练模型权重里1%的保矩阵QRA,再把权重量化到4bit,进一步省75%的内存,这就是Encelos内存优化的核心思路。", "tokens": [50365, 19683, 13876, 41646, 2257, 34592, 21596, 11, 43, 1607, 81, 14003, 7422, 255, 10115, 225, 41908, 39823, 4422, 225, 12624, 15759, 16, 4, 1546, 24302, 5881, 102, 10034, 113, 48, 3750, 11, 8623, 16075, 4422, 225, 12624, 26748, 23756, 4511, 19, 5260, 11, 36700, 2257, 31429, 2862, 223, 11901, 4, 1546, 34742, 39781, 11, 5562, 5620, 16257, 4933, 329, 34742, 39781, 7384, 246, 23756, 1546, 8987, 116, 7945, 8870, 24658, 1543, 50792], "temperature": 0, "avg_logprob": -0.08514818268506516, "compression_ratio": 1.226932668329177, "no_speech_prob": 1.4071750709110376e-11}, {"id": 31, "seek": 16204, "start": 170.72, "end": 179.06, "text": "除了训练,它的对话能力也不含糊,本地模型也能打工具,工具调用准确率,提升了50%,还带自愈机制,调用失败,会自动检测,修复。", "tokens": [50799, 32999, 2289, 7422, 255, 10115, 225, 11, 45224, 8713, 21596, 8225, 13486, 6404, 1960, 2392, 104, 15239, 232, 11, 8802, 10928, 41908, 39823, 6404, 8225, 12467, 23323, 39806, 11, 23323, 39806, 8897, 225, 9254, 6336, 228, 38114, 106, 44866, 11, 20949, 41670, 2289, 2803, 8923, 14852, 4845, 99, 9722, 4820, 230, 37960, 25491, 11, 8897, 225, 9254, 28564, 18464, 98, 11, 12949, 9722, 34961, 16407, 222, 11038, 233, 11, 7792, 106, 1787, 235, 1543, 51216], "temperature": 0, "avg_logprob": -0.08514818268506516, "compression_ratio": 1.226932668329177, "no_speech_prob": 1.4071750709110376e-11}, {"id": 32, "seek": 17906, "start": 179.06, "end": 194.8, "text": "从4,同时支持并行对话,几个任务同时聊,它还支持安全的代码执行,在沙箱里,运行Bass和Python,调用更稳,防止死循环,防止代码串话,权限这块也做得细,模型访问文件,联网执行命令都要经过你确认可以完全在安全沙箱里跑,也可以授权直接改文件。", "tokens": [50365, 35630, 19, 11, 13089, 15729, 35488, 3509, 114, 8082, 8713, 21596, 11, 6336, 254, 7549, 26443, 5087, 94, 13089, 15729, 40096, 11, 11284, 14852, 35488, 45466, 1546, 19105, 23230, 223, 3416, 100, 8082, 11, 3581, 3308, 247, 11249, 109, 15759, 11, 3316, 238, 8082, 33, 640, 12565, 47, 88, 11943, 11, 8897, 225, 9254, 19002, 10415, 111, 11, 35863, 30438, 16989, 2172, 103, 8051, 107, 11, 35863, 30438, 19105, 23230, 223, 940, 110, 21596, 11, 4422, 225, 43446, 5562, 47734, 6404, 10907, 5916, 10115, 228, 11, 41908, 39823, 7422, 123, 22064, 17174, 20485, 11, 8171, 242, 16469, 239, 3416, 100, 8082, 25236, 49061, 7182, 4275, 30276, 16866, 2166, 38114, 106, 7422, 97, 6723, 37100, 3581, 45466, 3308, 247, 11249, 109, 15759, 32585, 11, 47935, 6900, 230, 4422, 225, 43297, 34490, 17174, 20485, 1543, 51152], "temperature": 0, "avg_logprob": -0.061365961493932614, "compression_ratio": 1.3056872037914693, "no_speech_prob": 1.2173328317599541e-11}, {"id": 33, "seek": 19480, "start": 194.8, "end": 199.60000000000002, "text": "另外还有不限量的私人网络搜索,和一个深度研究模式,先规划,再查员,最后成文。", "tokens": [50365, 26202, 35091, 1960, 43446, 26748, 1546, 20083, 4035, 16469, 239, 10115, 250, 162, 14987, 7732, 95, 11, 12565, 20182, 24043, 13127, 23230, 242, 44704, 41908, 27584, 11, 10108, 6758, 226, 2437, 240, 11, 8623, 42623, 3606, 246, 11, 8661, 13547, 11336, 17174, 1543, 50605], "temperature": 0, "avg_logprob": -0.09917775711210647, "compression_ratio": 1.3517060367454068, "no_speech_prob": 1.4465958812770463e-11}, {"id": 34, "seek": 19480, "start": 199.8, "end": 208.68, "text": "引用都有出处,光自己对话还不够,它还能让你的AI agent直接用上本地模型,一条命令就能接入主流agent工具,模型随便换,数据不出门,用量不花钱。", "tokens": [50615, 41301, 9254, 48121, 7781, 1787, 226, 11, 20690, 17645, 8713, 21596, 14852, 1960, 1787, 253, 11, 11284, 14852, 8225, 33650, 18961, 48698, 9461, 43297, 9254, 5708, 8802, 10928, 41908, 39823, 11, 2257, 48837, 25236, 49061, 3111, 8225, 14468, 14028, 13557, 27854, 559, 317, 23323, 39806, 11, 41908, 39823, 10673, 237, 27364, 26075, 95, 11, 33188, 26075, 106, 1960, 7781, 8259, 101, 11, 9254, 26748, 1960, 20127, 39623, 1543, 51059], "temperature": 0, "avg_logprob": -0.09917775711210647, "compression_ratio": 1.3517060367454068, "no_speech_prob": 1.4465958812770463e-11}, {"id": 35, "seek": 19480, "start": 208.94, "end": 223.32000000000002, "text": "在终端里输入,onslots start,clawed,就能接入clawed code,输入onslots start codex,接入openai codex,还有hermes,openclaw,opencode,分别对应hermes agent,openclaw和opencode,这么做的好处,其实就三个。", "tokens": [51072, 3581, 10115, 230, 11957, 107, 15759, 9830, 241, 14028, 11, 892, 75, 1971, 722, 11, 66, 5901, 292, 11, 3111, 8225, 14468, 14028, 66, 5901, 292, 3089, 11, 9830, 241, 14028, 892, 75, 1971, 722, 3089, 87, 11, 14468, 14028, 15752, 1301, 3089, 87, 11, 35091, 511, 5814, 11, 15752, 66, 5901, 11, 15752, 22332, 11, 6627, 18453, 8713, 44297, 511, 5814, 9461, 11, 15752, 66, 5901, 12565, 15752, 22332, 11, 31889, 10907, 1546, 2131, 1787, 226, 11, 41646, 3111, 10960, 7549, 1543, 51791], "temperature": 0, "avg_logprob": -0.09917775711210647, "compression_ratio": 1.3517060367454068, "no_speech_prob": 1.4465958812770463e-11}, {"id": 36, "seek": 22332, "start": 223.32, "end": 227.72, "text": "第一,模型随便换,本地quian,deepseek,glm,想用哪个换哪个,随时切换。", "tokens": [50365, 18049, 11, 41908, 39823, 10673, 237, 27364, 26075, 95, 11, 8802, 10928, 358, 952, 11, 38422, 405, 916, 11, 70, 75, 76, 11, 7093, 9254, 17028, 7549, 26075, 95, 17028, 7549, 11, 10673, 237, 15729, 23632, 26075, 95, 1543, 50585], "temperature": 0, "avg_logprob": -0.06244991285383968, "compression_ratio": 1.3181818181818181, "no_speech_prob": 1.5148634083250734e-11}, {"id": 37, "seek": 22332, "start": 227.84, "end": 231.92, "text": "第二,代码不外泄,项目代码只在本地处理,再也不怕敏感,代码传云端。", "tokens": [50591, 19693, 11, 19105, 23230, 223, 1960, 12022, 6847, 226, 11, 10178, 117, 11386, 19105, 23230, 223, 14003, 3581, 8802, 10928, 1787, 226, 13876, 11, 8623, 6404, 1960, 21164, 7017, 237, 9709, 11, 19105, 23230, 223, 7384, 254, 1369, 239, 11957, 107, 1543, 50795], "temperature": 0, "avg_logprob": -0.06244991285383968, "compression_ratio": 1.3181818181818181, "no_speech_prob": 1.5148634083250734e-11}, {"id": 38, "seek": 22332, "start": 232.14, "end": 235.56, "text": "第三,用量不花钱,本地推理免费随便用,跑满也不心疼。", "tokens": [50806, 35878, 11, 9254, 26748, 1960, 20127, 39623, 11, 8802, 10928, 33597, 13876, 2347, 235, 18464, 117, 10673, 237, 27364, 9254, 11, 32585, 15868, 94, 6404, 1960, 7945, 18140, 120, 1543, 50977], "temperature": 0, "avg_logprob": -0.06244991285383968, "compression_ratio": 1.3181818181818181, "no_speech_prob": 1.5148634083250734e-11}, {"id": 39, "seek": 22332, "start": 235.79999999999998, "end": 245.78, "text": "模型够不够新,也是大家关心的事,它的生态非常丰富,支持500多个模型的推理与训练,文本,视觉,语音,嵌入,图像视频扩散,全类型覆盖。", "tokens": [50989, 41908, 39823, 1787, 253, 1960, 1787, 253, 12560, 11, 22021, 6868, 28053, 7945, 1546, 6973, 11, 45224, 8244, 3757, 223, 14392, 940, 108, 47564, 11, 35488, 7526, 6392, 7549, 41908, 39823, 1546, 33597, 13876, 940, 236, 7422, 255, 10115, 225, 11, 17174, 8802, 11, 40656, 24447, 11, 5233, 255, 18034, 11, 161, 113, 234, 14028, 11, 3919, 122, 12760, 40656, 39752, 3416, 102, 7017, 96, 11, 11319, 22113, 119, 39823, 2888, 228, 5419, 244, 1543, 51488], "temperature": 0, "avg_logprob": -0.06244991285383968, "compression_ratio": 1.3181818181818181, "no_speech_prob": 1.5148634083250734e-11}, {"id": 40, "seek": 24578, "start": 245.78, "end": 250.58, "text": "团队和多家模型厂商深度合作,做到当天首发,当天支持,叫DayZero。", "tokens": [50365, 3919, 95, 10034, 253, 12565, 6392, 5155, 41908, 39823, 5014, 224, 45581, 24043, 13127, 14245, 11914, 11, 10907, 4511, 16233, 6135, 25444, 28926, 11, 16233, 6135, 35488, 11, 19855, 38495, 57, 2032, 1543, 50605], "temperature": 0, "avg_logprob": -0.11600848782447076, "compression_ratio": 1.1503267973856208, "no_speech_prob": 1.439822133048052e-11}, {"id": 41, "seek": 24578, "start": 250.88, "end": 257.22, "text": "简单说,新模型当天发布,当天就能跑,能训不用等适配,第一时间体验最新大模型,最新的旗舰模型都能跑。", "tokens": [50620, 11249, 222, 47446, 8090, 11, 12560, 41908, 39823, 16233, 6135, 28926, 34688, 11, 16233, 6135, 3111, 8225, 32585, 11, 8225, 7422, 255, 24384, 10187, 2215, 224, 38846, 11, 18049, 44848, 29485, 49657, 234, 8661, 12560, 3582, 41908, 39823, 11, 8661, 12560, 1546, 4479, 245, 10256, 108, 41908, 39823, 7182, 8225, 32585, 1543, 50937], "temperature": 0, "avg_logprob": -0.11600848782447076, "compression_ratio": 1.1503267973856208, "no_speech_prob": 1.439822133048052e-11}, {"id": 42, "seek": 24578, "start": 257.6, "end": 267.12, "text": "比如quian 3.8 27b,kimik3,deepseek v4,glm5.3,minimax h3,还有museglimmer,jammer4,本地已有的GGUF模型,也能自动识别。", "tokens": [50956, 36757, 358, 952, 805, 13, 23, 7634, 65, 11, 25112, 1035, 18, 11, 38422, 405, 916, 371, 19, 11, 70, 75, 76, 20, 13, 18, 11, 2367, 332, 2797, 276, 18, 11, 35091, 76, 438, 70, 4197, 936, 11, 24267, 936, 19, 11, 8802, 10928, 15701, 2412, 1546, 27561, 52, 37, 41908, 39823, 11, 6404, 8225, 9722, 34961, 5233, 228, 18453, 1543, 51432], "temperature": 0, "avg_logprob": -0.11600848782447076, "compression_ratio": 1.1503267973856208, "no_speech_prob": 1.439822133048052e-11}, {"id": 43, "seek": 26712, "start": 267.12, "end": 269.7, "text": "不用重新下载,也可以手动指定自定义目录。", "tokens": [50365, 24384, 12624, 12560, 4438, 17819, 121, 11, 47935, 11389, 34961, 25922, 12088, 9722, 12088, 2930, 231, 11386, 7391, 243, 1543, 50494], "temperature": 0, "avg_logprob": -0.11728545903090404, "compression_ratio": 1.2150259067357514, "no_speech_prob": 1.5024790439577274e-11}, {"id": 44, "seek": 26712, "start": 269.78000000000003, "end": 273.64, "text": "它还不只是文本模型,图像,视频,语音,通通能训,能生成。", "tokens": [50498, 11284, 14852, 1960, 36859, 17174, 8802, 41908, 39823, 11, 3919, 122, 12760, 11, 40656, 39752, 11, 5233, 255, 18034, 11, 19550, 19550, 8225, 7422, 255, 11, 8225, 8244, 11336, 1543, 50691], "temperature": 0, "avg_logprob": -0.11728545903090404, "compression_ratio": 1.2150259067357514, "no_speech_prob": 1.5024790439577274e-11}, {"id": 45, "seek": 26712, "start": 273.76, "end": 278.38, "text": "图像视频方面,支持Flux,Z-Image,LTX-1,这些模型还能训练自己的扩散。", "tokens": [50697, 3919, 122, 12760, 40656, 39752, 39604, 11, 35488, 25680, 2449, 11, 57, 12, 31128, 609, 11, 43, 51, 55, 12, 16, 11, 5562, 13824, 41908, 39823, 14852, 8225, 7422, 255, 10115, 225, 17645, 1546, 3416, 102, 7017, 96, 1543, 50928], "temperature": 0, "avg_logprob": -0.11728545903090404, "compression_ratio": 1.2150259067357514, "no_speech_prob": 1.5024790439577274e-11}, {"id": 46, "seek": 26712, "start": 278.62, "end": 288.68, "text": "Laura,比如SDXL,FLUX 2.0,Q1,Image,生成速度也很快,用minimax h3,生成一段960x544,124帧,8部的视频从70多秒,压缩到13秒。", "tokens": [50940, 5478, 2991, 11, 36757, 23969, 55, 43, 11, 31455, 52, 55, 568, 13, 15, 11, 48, 16, 11, 31128, 609, 11, 8244, 11336, 31217, 13127, 6404, 4563, 10251, 11, 9254, 2367, 332, 2797, 276, 18, 11, 8244, 11336, 2257, 28427, 24, 4550, 87, 20, 13912, 11, 4762, 19, 4845, 100, 11, 23, 13470, 1546, 40656, 39752, 35630, 5867, 6392, 20956, 11, 5014, 233, 38109, 102, 4511, 7668, 20956, 1543, 51443], "temperature": 0, "avg_logprob": -0.11728545903090404, "compression_ratio": 1.2150259067357514, "no_speech_prob": 1.5024790439577274e-11}, {"id": 47, "seek": 26712, "start": 288.68, "end": 293.22, "text": "语音方面,文字转语音,语音转文字,Whisper,Q1,3ASR,全部本地处理。", "tokens": [51443, 5233, 255, 18034, 39604, 11, 17174, 22381, 17819, 105, 5233, 255, 18034, 11, 5233, 255, 18034, 17819, 105, 17174, 22381, 11, 2471, 271, 610, 11, 48, 16, 11, 18, 3160, 49, 11, 38714, 8802, 10928, 1787, 226, 13876, 1543, 51670], "temperature": 0, "avg_logprob": -0.11728545903090404, "compression_ratio": 1.2150259067357514, "no_speech_prob": 1.5024790439577274e-11}, {"id": 48, "seek": 29322, "start": 293.22, "end": 298.62, "text": "一份工具,覆盖文本图像视频,语音四条生成线,训练生成,微调,转写都在同一个界面完成。", "tokens": [50365, 2257, 36266, 23323, 39806, 11, 2888, 228, 5419, 244, 17174, 8802, 3919, 122, 12760, 40656, 39752, 11, 5233, 255, 18034, 19425, 48837, 8244, 11336, 16853, 123, 11, 7422, 255, 10115, 225, 8244, 11336, 11, 39152, 8897, 225, 11, 17819, 105, 5676, 247, 7182, 3581, 13089, 20182, 21185, 8833, 41509, 1543, 50635], "temperature": 0, "avg_logprob": -0.062009490658884094, "compression_ratio": 1.2468193384223918, "no_speech_prob": 1.3300565510077078e-11}, {"id": 49, "seek": 29322, "start": 298.8, "end": 301.5, "text": "训练完的模型,怎么部署,它给了你三种方式。", "tokens": [50644, 7422, 255, 10115, 225, 14128, 1546, 41908, 39823, 11, 15282, 13470, 16469, 110, 11, 11284, 23197, 2289, 2166, 10960, 39810, 9249, 27584, 1543, 50779], "temperature": 0, "avg_logprob": -0.062009490658884094, "compression_ratio": 1.2468193384223918, "no_speech_prob": 1.3300565510077078e-11}, {"id": 50, "seek": 29322, "start": 301.66, "end": 304.34000000000003, "text": "第一种,局域网访问,直接开启LNN模式。", "tokens": [50787, 18049, 39810, 11, 34703, 16262, 253, 16469, 239, 7422, 123, 22064, 11, 43297, 18937, 2392, 107, 43, 45, 45, 41908, 27584, 1543, 50921], "temperature": 0, "avg_logprob": -0.062009490658884094, "compression_ratio": 1.2468193384223918, "no_speech_prob": 1.3300565510077078e-11}, {"id": 51, "seek": 29322, "start": 304.52000000000004, "end": 308.3, "text": "局域网内任意设备都能访问,流量不出网,延迟低,断网也能用。", "tokens": [50930, 34703, 16262, 253, 16469, 239, 34742, 26443, 9042, 7422, 122, 1787, 229, 7182, 8225, 7422, 123, 22064, 11, 27854, 26748, 1960, 7781, 16469, 239, 11, 17527, 114, 3316, 253, 41377, 11, 4307, 255, 16469, 239, 6404, 8225, 9254, 1543, 51119], "temperature": 0, "avg_logprob": -0.062009490658884094, "compression_ratio": 1.2468193384223918, "no_speech_prob": 1.3300565510077078e-11}, {"id": 52, "seek": 29322, "start": 308.48, "end": 314.90000000000003, "text": "第二种,公网HTPPS,用免费的Cloud Flare隧道,把本地模型安全暴露到公网手机上随时能访问。", "tokens": [51128, 19693, 39810, 11, 13545, 16469, 239, 39, 16804, 6273, 11, 9254, 2347, 235, 18464, 117, 1546, 32787, 3235, 543, 10673, 100, 6025, 11, 16075, 8802, 10928, 41908, 39823, 45466, 23826, 112, 18594, 110, 4511, 13545, 16469, 239, 11389, 37960, 5708, 10673, 237, 15729, 8225, 7422, 123, 22064, 1543, 51449], "temperature": 0, "avg_logprob": -0.062009490658884094, "compression_ratio": 1.2468193384223918, "no_speech_prob": 1.3300565510077078e-11}, {"id": 53, "seek": 31490, "start": 314.9, "end": 319.91999999999996, "text": "第三种,OpenAI兼容API,提供标准接口,任何支持这个协议的应用都能直接接入。", "tokens": [50365, 35878, 39810, 11, 45569, 48698, 2347, 120, 25750, 4715, 40, 11, 20949, 3254, 249, 162, 3921, 6336, 228, 14468, 18144, 11, 26443, 11906, 35488, 15368, 5322, 237, 7422, 106, 1546, 44297, 9254, 7182, 8225, 43297, 14468, 14028, 1543, 50616], "temperature": 0, "avg_logprob": -0.09813620338977223, "compression_ratio": 1.1985645933014355, "no_speech_prob": 1.1743594811874924e-11}, {"id": 54, "seek": 31490, "start": 320.2, "end": 326.03999999999996, "text": "你也可以用Uncelot Run命令,启动模型服务,指定模型和推理参数,比如温度,采样方式这些就能对外提供接口。", "tokens": [50630, 2166, 47935, 9254, 12405, 4933, 310, 8950, 25236, 49061, 11, 2392, 107, 34961, 41908, 39823, 27408, 5087, 94, 11, 25922, 12088, 41908, 39823, 12565, 33597, 13876, 2129, 224, 33188, 11, 36757, 9592, 102, 13127, 11, 5873, 229, 14496, 9249, 27584, 5562, 13824, 3111, 8225, 8713, 12022, 20949, 3254, 249, 14468, 18144, 1543, 50922], "temperature": 0, "avg_logprob": -0.09813620338977223, "compression_ratio": 1.1985645933014355, "no_speech_prob": 1.1743594811874924e-11}, {"id": 55, "seek": 31490, "start": 326.23999999999995, "end": 333.62, "text": "而且还能连接云端Provider,OpenAI,Anthropic,Alama,LamaCPP,VLLM,本地云端同一个界面来回切换。", "tokens": [50932, 22942, 14852, 8225, 3316, 252, 14468, 1369, 239, 11957, 107, 12681, 85, 1438, 11, 45569, 48698, 11, 7828, 14222, 299, 11, 9171, 2404, 11, 43, 2404, 34, 17755, 11, 53, 24010, 44, 11, 8802, 10928, 1369, 239, 11957, 107, 13089, 20182, 21185, 8833, 6912, 8350, 23632, 26075, 95, 1543, 51301], "temperature": 0, "avg_logprob": -0.09813620338977223, "compression_ratio": 1.1985645933014355, "no_speech_prob": 1.1743594811874924e-11}, {"id": 56, "seek": 31490, "start": 333.62, "end": 335.28, "text": "最后,隐私和兼容性。", "tokens": [51301, 8661, 13547, 11, 10673, 238, 20083, 12565, 2347, 120, 25750, 21686, 1543, 51384], "temperature": 0, "avg_logprob": -0.09813620338977223, "compression_ratio": 1.1985645933014355, "no_speech_prob": 1.1743594811874924e-11}, {"id": 57, "seek": 31490, "start": 335.47999999999996, "end": 337.23999999999995, "text": "很多人特别关心,这里都讲清楚。", "tokens": [51394, 20778, 4035, 17682, 18453, 28053, 7945, 11, 35102, 7182, 39255, 38433, 1543, 51482], "temperature": 0, "avg_logprob": -0.09813620338977223, "compression_ratio": 1.1985645933014355, "no_speech_prob": 1.1743594811874924e-11}, {"id": 58, "seek": 31490, "start": 337.47999999999996, "end": 341.28, "text": "隐私方面,零摇测,不上传任何使用数据,完全离线,也能正常跑。", "tokens": [51494, 10673, 238, 20083, 39604, 11, 6306, 114, 34783, 229, 11038, 233, 11, 1960, 5708, 7384, 254, 26443, 11906, 22982, 9254, 33188, 26075, 106, 11, 37100, 18614, 119, 16853, 123, 11, 6404, 8225, 15789, 11279, 32585, 1543, 51684], "temperature": 0, "avg_logprob": -0.09813620338977223, "compression_ratio": 1.1985645933014355, "no_speech_prob": 1.1743594811874924e-11}, {"id": 59, "seek": 34128, "start": 341.28, "end": 345.11999999999995, "text": "只会检测本机GPU类型来适配运行,硬件要求也不高。", "tokens": [50365, 14003, 12949, 16407, 222, 11038, 233, 8802, 37960, 38, 8115, 22113, 119, 39823, 6912, 2215, 224, 38846, 3316, 238, 8082, 11, 38114, 105, 20485, 4275, 32718, 6404, 1960, 12979, 1543, 50557], "temperature": 0, "avg_logprob": -0.0955870870197789, "compression_ratio": 1.17312661498708, "no_speech_prob": 1.3319624916907635e-11}, {"id": 60, "seek": 34128, "start": 345.44, "end": 349.59999999999997, "text": "Windows 10或11,Mac OS 12以上,Ubuntu 20.04以上都行。", "tokens": [50573, 45813, 1509, 1266, 19780, 5348, 11, 34355, 12731, 2272, 29497, 11, 52, 65, 45605, 945, 13, 14565, 29497, 7182, 8082, 1543, 50781], "temperature": 0, "avg_logprob": -0.0955870870197789, "compression_ratio": 1.17312661498708, "no_speech_prob": 1.3319624916907635e-11}, {"id": 61, "seek": 34128, "start": 350.02, "end": 354.08, "text": "VDI,AMD,Intel,显卡加Mac都能训练,存CPU也能用。", "tokens": [50802, 53, 3085, 11, 2865, 35, 11, 25597, 338, 11, 1431, 122, 32681, 9990, 34355, 7182, 8225, 7422, 255, 10115, 225, 11, 39781, 34, 8115, 6404, 8225, 9254, 1543, 51005], "temperature": 0, "avg_logprob": -0.0955870870197789, "compression_ratio": 1.17312661498708, "no_speech_prob": 1.3319624916907635e-11}, {"id": 62, "seek": 34128, "start": 354.4, "end": 355.73999999999995, "text": "对话和数据处理,没问题。", "tokens": [51021, 8713, 21596, 12565, 33188, 26075, 106, 1787, 226, 13876, 11, 10062, 34069, 1543, 51088], "temperature": 0, "avg_logprob": -0.0955870870197789, "compression_ratio": 1.17312661498708, "no_speech_prob": 1.3319624916907635e-11}, {"id": 63, "seek": 34128, "start": 356.02, "end": 359.61999999999995, "text": "几个常见问题也顺带回答,已经下载的模型能不能用,能自动识别?", "tokens": [51102, 6336, 254, 7549, 11279, 23813, 34069, 6404, 10178, 118, 4845, 99, 8350, 28223, 11, 49161, 4438, 17819, 121, 1546, 41908, 39823, 8225, 28590, 9254, 11, 8225, 9722, 34961, 5233, 228, 18453, 30, 51282], "temperature": 0, "avg_logprob": -0.0955870870197789, "compression_ratio": 1.17312661498708, "no_speech_prob": 1.3319624916907635e-11}, {"id": 64, "seek": 34128, "start": 359.85999999999996, "end": 362.44, "text": "必须高端显卡吗?不用CPU,Mac都能跑。", "tokens": [51294, 28531, 10178, 119, 12979, 11957, 107, 1431, 122, 32681, 14769, 30, 24384, 34, 8115, 11, 34355, 7182, 8225, 32585, 1543, 51423], "temperature": 0, "avg_logprob": -0.0955870870197789, "compression_ratio": 1.17312661498708, "no_speech_prob": 1.3319624916907635e-11}, {"id": 65, "seek": 34128, "start": 362.71999999999997, "end": 364.23999999999995, "text": "支持OpenAI兼容API吗?", "tokens": [51437, 35488, 45569, 48698, 2347, 120, 25750, 4715, 40, 14769, 30, 51513], "temperature": 0, "avg_logprob": -0.0955870870197789, "compression_ratio": 1.17312661498708, "no_speech_prob": 1.3319624916907635e-11}, {"id": 66, "seek": 34128, "start": 364.53999999999996, "end": 367.91999999999996, "text": "支持推理,偶尔曼,把联网搜索和代码执行关掉,速度立刻回来。", "tokens": [51528, 35488, 33597, 13876, 11, 7437, 114, 1530, 242, 9531, 120, 11, 16075, 8171, 242, 16469, 239, 162, 14987, 7732, 95, 12565, 19105, 23230, 223, 3416, 100, 8082, 28053, 29327, 11, 31217, 13127, 24409, 45500, 8350, 6912, 1543, 51697], "temperature": 0, "avg_logprob": -0.0955870870197789, "compression_ratio": 1.17312661498708, "no_speech_prob": 1.3319624916907635e-11}, {"id": 67, "seek": 36792, "start": 367.92, "end": 371.56, "text": "说到这,Unclosed Desk的OPPO的核心价值已经很清楚了,一句话,", "tokens": [50365, 8090, 4511, 5562, 11, 12405, 3474, 1744, 3885, 74, 1546, 12059, 34885, 1546, 8987, 116, 7945, 1550, 115, 40242, 49161, 4563, 38433, 2289, 11, 2257, 34592, 21596, 11, 50547], "temperature": 0, "avg_logprob": -0.14443581414329631, "compression_ratio": 1.2532981530343008, "no_speech_prob": 1.4261815686755774e-11}, {"id": 68, "seek": 36792, "start": 371.94, "end": 375.72, "text": "总结它把本地训练大模型从极客专属变成了人人都能上手。", "tokens": [50566, 33440, 45641, 11284, 16075, 8802, 10928, 7422, 255, 10115, 225, 3582, 41908, 39823, 35630, 7360, 223, 32316, 940, 241, 9636, 252, 2129, 246, 11336, 2289, 4035, 4035, 7182, 8225, 5708, 11389, 1543, 50755], "temperature": 0, "avg_logprob": -0.14443581414329631, "compression_ratio": 1.2532981530343008, "no_speech_prob": 1.4261815686755774e-11}, {"id": 69, "seek": 36792, "start": 375.98, "end": 379.86, "text": "它的核心价值就三个词,本地化,免代码,全类型模型支持,", "tokens": [50768, 45224, 8987, 116, 7945, 1550, 115, 40242, 3111, 10960, 7549, 5233, 235, 11, 8802, 10928, 23756, 11, 2347, 235, 19105, 23230, 223, 11, 11319, 22113, 119, 39823, 41908, 39823, 35488, 11, 50962], "temperature": 0, "avg_logprob": -0.14443581414329631, "compression_ratio": 1.2532981530343008, "no_speech_prob": 1.4261815686755774e-11}, {"id": 70, "seek": 36792, "start": 380.20000000000005, "end": 382.34000000000003, "text": "这也是桌面端本地训练大模型的新选择。", "tokens": [50979, 5562, 22021, 162, 5475, 8833, 11957, 107, 8802, 10928, 7422, 255, 10115, 225, 3582, 41908, 39823, 1546, 12560, 2215, 231, 6852, 102, 1543, 51086], "temperature": 0, "avg_logprob": -0.14443581414329631, "compression_ratio": 1.2532981530343008, "no_speech_prob": 1.4261815686755774e-11}, {"id": 71, "seek": 36792, "start": 382.8, "end": 385.20000000000005, "text": "回顾一下,无代码微调,丢文件就能训。", "tokens": [51109, 8350, 10178, 122, 8861, 11, 31364, 19105, 23230, 223, 39152, 8897, 225, 11, 940, 95, 17174, 20485, 3111, 8225, 7422, 255, 1543, 51229], "temperature": 0, "avg_logprob": -0.14443581414329631, "compression_ratio": 1.2532981530343008, "no_speech_prob": 1.4261815686755774e-11}, {"id": 72, "seek": 36792, "start": 385.66, "end": 388.04, "text": "两倍提速70%,神险存,精度无损。", "tokens": [51252, 36257, 35477, 20949, 31217, 5867, 8923, 20432, 8842, 102, 39781, 11, 34910, 13127, 31364, 26075, 253, 1543, 51371], "temperature": 0, "avg_logprob": -0.14443581414329631, "compression_ratio": 1.2532981530343008, "no_speech_prob": 1.4261815686755774e-11}, {"id": 73, "seek": 36792, "start": 388.48, "end": 390.02000000000004, "text": "一行命令,接入AI agent。", "tokens": [51393, 2257, 8082, 25236, 49061, 11, 14468, 14028, 48698, 9461, 1543, 51470], "temperature": 0, "avg_logprob": -0.14443581414329631, "compression_ratio": 1.2532981530343008, "no_speech_prob": 1.4261815686755774e-11}, {"id": 74, "seek": 36792, "start": 390.48, "end": 392.56, "text": "局域网,军网,API,全能部署。", "tokens": [51493, 34703, 16262, 253, 16469, 239, 11, 5676, 249, 16469, 239, 11, 4715, 40, 11, 11319, 8225, 13470, 16469, 110, 1543, 51597], "temperature": 0, "avg_logprob": -0.14443581414329631, "compression_ratio": 1.2532981530343008, "no_speech_prob": 1.4261815686755774e-11}, {"id": 75, "seek": 39256, "start": 392.56, "end": 394.88, "text": "零瑶测,全力线,数据不出门,适合谁用。", "tokens": [50365, 6306, 114, 37955, 114, 11038, 233, 11, 11319, 13486, 16853, 123, 11, 33188, 26075, 106, 1960, 7781, 8259, 101, 11, 2215, 224, 14245, 33556, 9254, 1543, 50481], "temperature": 0, "avg_logprob": -0.09788360595703124, "compression_ratio": 1.1764705882352942, "no_speech_prob": 1.0568224247109459e-11}, {"id": 76, "seek": 39256, "start": 395.22, "end": 398.04, "text": "想本地微调模型的开发者注重数据隐私的企业和个人,", "tokens": [50498, 7093, 8802, 10928, 39152, 8897, 225, 41908, 39823, 1546, 18937, 28926, 12444, 26432, 12624, 33188, 26075, 106, 10673, 238, 20083, 1546, 7384, 223, 940, 248, 12565, 7549, 4035, 11, 50639], "temperature": 0, "avg_logprob": -0.09788360595703124, "compression_ratio": 1.1764705882352942, "no_speech_prob": 1.0568224247109459e-11}, {"id": 77, "seek": 39256, "start": 398.28000000000003, "end": 401.46, "text": "想省云费用的用户,还有刚接触大模型的新手都值得试试。", "tokens": [50651, 7093, 2862, 223, 1369, 239, 18464, 117, 9254, 1546, 9254, 1486, 115, 11, 35091, 49160, 14468, 6758, 99, 3582, 41908, 39823, 1546, 12560, 11389, 7182, 40242, 5916, 5233, 243, 5233, 243, 1543, 50810], "temperature": 0, "avg_logprob": -0.09788360595703124, "compression_ratio": 1.1764705882352942, "no_speech_prob": 1.0568224247109459e-11}, {"id": 78, "seek": 39256, "start": 401.78000000000003, "end": 404.88, "text": "感谢观看,我是大叔大,后续会持续更新,下期见。", "tokens": [50826, 9709, 11340, 42350, 4200, 11, 15914, 3582, 39936, 3582, 11, 13547, 10115, 255, 12949, 17694, 10115, 255, 19002, 12560, 11, 4438, 16786, 23813, 1543, 50981], "temperature": 0, "avg_logprob": -0.09788360595703124, "compression_ratio": 1.1764705882352942, "no_speech_prob": 1.0568224247109459e-11}], "language": "zh"}