{ "text": "今天DeepSight 4.1 Flash正式上线了。昨天内测的时候,我只来得及测完基本能力。本期给大家把剩下的测试全部补齐。官网已经把原来的快速、专家、实图三个模式合并升级成一个对话框,背后跑的就是4.1 Flash。我也验证了API同样已经升级到了4.1。另外,还有一个好消息和一个坏消息。好消息是Flash上线就立马降价。我测试了不同的场景。综合来看,降幅在百分之一,\n百分之二十到百分之四十五。坏消息是,四天后,我最喜欢的Pro版就要下线了。官方公告只有屏幕上这一小段文字。14号以后,发往Pro模型的请求会自动路由到新的4.1 Flash,并按Flash单价计费。给出的理由是,Flash在性能、费用和速度上已经全部超过了Pro。真的是这样吗?我会给你一一分解,有一肚子化不吐不快。官方公布了新的Flash模型架构。\n4.1 Flash的总参数量提升了接近一倍。而且输入激活参数和输出激活参数量是不同的。输入参数降了百分之四十,输出则增加了百分之二十。这是一种全新的模型结构。\n有意思的是,模型更大了,API价格反而更低了,其中降幅最大的是缓存命中,直接降了百分之六十。至于降价的原因,屏幕上这张官方图就很形象。\n这代新的Flash模型架构大幅降低了对KV缓存的硬件压力。图片对比了每一代Deepseq每一个Token所需的缓存,在短短两年内降到了千分之二。\n这也让它的内存和SSD需求分别降到了上一代模型的四分之一和八分之一。正是因为成本降下来了,才能给到用户实打实的福利。\n测试正式开始。跟上一代视觉实验版相比,在第二轮的测试里,中文写作和实图能力持平,前端视觉有一定的提升。\n但我不得不收回它能达到前端Flash模型天花板的话。办公任务依然稳如老狗,而全战任务跟视觉实验版刚好打平。\n具体的分数和计费我都列在屏幕上了。这些测试往期视频讲了太多次,今天只挑两个给大家看看。\n先看前端能力,这次还是用外滩的昼夜风景来测试。\n这个版本比起上一代儿童涂鸦式的呈现已经强了很多,至少能分辨出厨房三件套了,昼夜光线的变化也比较连贯自然。\n但是,它仍然是非常简单的几何图案堆叠。\n另外有一点让我很惊讶,就是它消耗了比原来多一倍的token。\n为了让大家有个直观印象,我把这个测试项目里的冠军给大家看一下。\n这么一笔Deepseek的Flash模型的前端能力,还得加油啊。\n大家看看我的保留项目。\n这是一个复杂的长城任务,需要处理大约1G的原始素材,从零搭建数据库,构建整个小说阅读站的前后端代码。\n4.1,Flash的速度极度亮眼,21分钟就完成了,燃烧了3210万token。\n它的实现我仔细看了一遍,有三个明显的前端bug和一些其他小问题,大家可以看屏幕。\n这模型干活的速度可以说是快如闪电,但做事的脾气却非常中规中矩。\n它能很稳地实现你的要求,但绝不会多做任何需求之外的事情。\n跟上一代视觉实验版相比,这个项目两边的分数完全持平,所以我是有点惊讶的,因为按照官方的说法,理应有很大的提升。\n来聊聊今天我想吐槽的第一个问题,4.1,Flash,真能代替Pro吗?\n按照这个表格,Flash更擅长前端任务和Pro本身不支持的实图,而Pro在复杂任务上仍然稳稳地压过Flash。\n尤其是中文写作上的差距,直接让我痛失了最常用的写作模型。\n当然,对于每天用模型真正干活的人来说,速度也是非常重要的,等10分钟和等半小时,感受完全不一样。\n所以在我看来,Pro和Flash其实是非常互补,缺一不可的,快速出前端视觉用Flash,而创意写作和复杂任务用Pro保证出品质量。\n那为什么Deepseek还要强行把Pro下架?\n看到这个消息的时候,我真的是两眼一黑,而且并不是只有我一个人有这种体会。\n大模型社区里早就炒番了,少部分人是很开心的,因为他们很期待又快又有多模态的新款Flash,但另一派则是非常愤怒。\n因为Pro模型的世界知识,更好的创意写作,以及在复杂任务里的稳定性,就这么被硬生生剥夺了。\n而且这事只给了4天的时间做缓冲。\n按照正常的商业逻辑,这种级别的切换,不应该至少给一个月的牵移期吗?\n最后再多点我个人的观点,首先,恭喜Deepseek跑通了优秀的架构,首个多模态Flash上线,而且确实速度快效率高。\n但另一边,那只Pro大鲸鱼就这么被血藏了,这件事我心里确实过不去。\n作为一个每天都在跟大模型打交道的程序员,我已经把自动脚本里的Pro模型迁移到了其他开源模型供应商,但仍然觉得这个切换太过于仓促了。\n另外呢,平心而论,只靠Flash模型是撑不起一家大模型公司的未来的。\n所以,我个人非常非常希望能看到公告里说的4.1Pro尽快上线。\n如果我说的不对,欢迎在评论区指证,每次我都能在那儿看到特别精彩的发言,就等你来哦。", "language": "zh", "duration": 314.93225, "segments": [ { "id": 0, "seek": 0, "start": 0.0, "end": 29.98, "text": "今天DeepSight 4.1 Flash正式上线了。昨天内测的时候,我只来得及测完基本能力。本期给大家把剩下的测试全部补齐。官网已经把原来的快速、专家、实图三个模式合并升级成一个对话框,背后跑的就是4.1 Flash。我也验证了API同样已经升级到了4.1。另外,还有一个好消息和一个坏消息。好消息是Flash上线就立马降价。我测试了不同的场景。综合来看,降幅在百分之一,", "tokens": [ 50365, 12074, 11089, 595, 50, 397, 1017, 13, 16, 20232, 15789, 27584, 5708, 16853, 123, 2289, 1543, 47404, 6135, 34742, 11038, 233, 49873, 11, 1654, 14003, 6912, 5916, 25703, 11038, 233, 14128, 37946, 8225, 13486, 1543, 8802, 16786, 23197, 6868, 16075, 5935, 102, 4438, 1546, 11038, 233, 5233, 243, 38714, 9890, 98, 44628, 238, 1543, 31929, 16469, 239, 49161, 16075, 19683, 6912, 1546, 10251, 31217, 1231, 940, 241, 5155, 1231, 24726, 3919, 122, 10960, 7549, 41908, 27584, 14245, 3509, 114, 41670, 16853, 100, 11336, 20182, 8713, 21596, 21416, 228, 11, 46329, 13547, 32585, 1546, 5620, 19, 13, 16, 20232, 1543, 27761, 49657, 234, 5233, 223, 2289, 4715, 40, 13089, 14496, 49161, 41670, 16853, 100, 21381, 19, 13, 16, 1543, 26202, 11, 35091, 20182, 2131, 28837, 26460, 12565, 20182, 14872, 237, 28837, 26460, 1543, 2131, 28837, 26460, 1541, 25680, 1299, 5708, 16853, 123, 3111, 24409, 49970, 47421, 1550, 115, 1543, 1654, 11038, 233, 5233, 243, 2289, 47123, 1546, 50255, 50218, 1543, 10115, 120, 14245, 6912, 4200, 11, 47421, 3509, 227, 3581, 31906, 6627, 9574, 2257, 11, 51864 ], "temperature": 0, "avg_logprob": -0.0981914113076885, "compression_ratio": 1.2853260869565217, "no_speech_prob": 9.358173090612265e-12, "source": "whisper" }, { "id": 1, "seek": 2998, "start": 29.98, "end": 59.14, "text": "百分之二十到百分之四十五。坏消息是,四天后,我最喜欢的Pro版就要下线了。官方公告只有屏幕上这一小段文字。14号以后,发往Pro模型的请求会自动路由到新的4.1 Flash,并按Flash单价计费。给出的理由是,Flash在性能、费用和速度上已经全部超过了Pro。真的是这样吗?我会给你一一分解,有一肚子化不吐不快。官方公布了新的Flash模型架构。", "tokens": [ 50365, 31906, 6627, 9574, 11217, 20145, 4511, 31906, 6627, 9574, 19425, 20145, 21001, 1543, 14872, 237, 28837, 26460, 1541, 11, 19425, 6135, 13547, 11, 1654, 8661, 41606, 1546, 12681, 42096, 3111, 4275, 4438, 16853, 123, 2289, 1543, 31929, 9249, 13545, 16846, 35244, 9636, 237, 40211, 5708, 5562, 2257, 7322, 28427, 17174, 22381, 1543, 7271, 26987, 3588, 13547, 11, 28926, 29510, 12681, 41908, 39823, 1546, 27908, 32718, 12949, 9722, 34961, 24658, 23786, 4511, 12560, 1546, 19, 13, 16, 20232, 11, 3509, 114, 26613, 25680, 1299, 47446, 1550, 115, 7422, 94, 18464, 117, 1543, 23197, 7781, 1546, 13876, 23786, 1541, 11, 25680, 1299, 3581, 21686, 8225, 1231, 18464, 117, 9254, 12565, 31217, 13127, 5708, 49161, 38714, 19869, 16866, 2289, 12681, 1543, 45698, 21209, 14769, 30, 1654, 12949, 23197, 2166, 2257, 2257, 6627, 17278, 11, 32241, 14356, 248, 7626, 23756, 1960, 2392, 238, 1960, 10251, 1543, 31929, 9249, 13545, 34688, 2289, 12560, 1546, 25680, 1299, 41908, 39823, 7360, 114, 7360, 226, 1543, 51823 ], "temperature": 0, "avg_logprob": -0.04423998609001254, "compression_ratio": 1.26, "no_speech_prob": 1.4407564551122132e-11, "source": "whisper" }, { "id": 2, "seek": 5914, "start": 59.14, "end": 72.54, "text": "4.1 Flash的总参数量提升了接近一倍。而且输入激活参数和输出激活参数量是不同的。输入参数降了百分之四十,输出则增加了百分之二十。这是一种全新的模型结构。", "tokens": [ 50365, 19, 13, 16, 20232, 1546, 33440, 2129, 224, 33188, 26748, 20949, 41670, 2289, 14468, 17463, 2257, 35477, 1543, 22942, 9830, 241, 14028, 26373, 222, 25956, 2129, 224, 33188, 12565, 9830, 241, 7781, 26373, 222, 25956, 2129, 224, 33188, 26748, 1541, 47123, 1546, 1543, 9830, 241, 14028, 2129, 224, 33188, 47421, 2289, 31906, 6627, 9574, 19425, 20145, 11, 9830, 241, 7781, 2437, 247, 24228, 252, 9990, 2289, 31906, 6627, 9574, 11217, 20145, 1543, 27455, 2257, 39810, 11319, 12560, 1546, 41908, 39823, 45641, 7360, 226, 1543, 51035 ], "temperature": 0, "avg_logprob": -0.04313407702879472, "compression_ratio": 1.2259887005649717, "no_speech_prob": 1.5398324976212407e-11, "source": "whisper" }, { "id": 3, "seek": 7254, "start": 72.54, "end": 84.54, "text": "有意思的是,模型更大了,API价格反而更低了,其中降幅最大的是缓存命中,直接降了百分之六十。至于降价的原因,屏幕上这张官方图就很形象。", "tokens": [ 50365, 2412, 16697, 24620, 11, 41908, 39823, 19002, 3582, 2289, 11, 4715, 40, 1550, 115, 30921, 22138, 11070, 19002, 41377, 2289, 11, 9572, 5975, 47421, 3509, 227, 8661, 3582, 24620, 38109, 241, 39781, 25236, 5975, 11, 43297, 47421, 2289, 31906, 6627, 9574, 30566, 20145, 1543, 20844, 37732, 47421, 1550, 115, 1546, 42577, 11, 9636, 237, 40211, 5708, 5562, 44059, 31929, 9249, 3919, 122, 3111, 4563, 30900, 45007, 1543, 50965 ], "temperature": 0, "avg_logprob": -0.049378982731993766, "compression_ratio": 1.1011904761904763, "no_speech_prob": 1.6827365889593437e-11, "source": "whisper" }, { "id": 4, "seek": 8454, "start": 84.54, "end": 96.54, "text": "这代新的Flash模型架构大幅降低了对KV缓存的硬件压力。图片对比了每一代Deepseq每一个Token所需的缓存,在短短两年内降到了千分之二。", "tokens": [ 50365, 5562, 19105, 12560, 1546, 25680, 1299, 41908, 39823, 7360, 114, 7360, 226, 3582, 3509, 227, 47421, 41377, 2289, 8713, 42, 53, 38109, 241, 39781, 1546, 38114, 105, 20485, 5014, 233, 13486, 1543, 3919, 122, 16668, 8713, 11706, 2289, 23664, 2257, 19105, 11089, 595, 405, 80, 23664, 20182, 51, 8406, 5966, 32535, 1546, 38109, 241, 39781, 11, 3581, 5881, 255, 5881, 255, 36257, 5157, 34742, 47421, 21381, 20787, 6627, 9574, 11217, 1543, 50965 ], "temperature": 0, "avg_logprob": -0.07540067036946614, "compression_ratio": 1.0114942528735633, "no_speech_prob": 1.6804120595015348e-11, "source": "whisper" }, { "id": 5, "seek": 9654, "start": 96.54, "end": 107.42, "text": "这也让它的内存和SSD需求分别降到了上一代模型的四分之一和八分之一。正是因为成本降下来了,才能给到用户实打实的福利。", "tokens": [ 50365, 5562, 6404, 33650, 45224, 34742, 39781, 12565, 21929, 35, 32535, 32718, 6627, 18453, 47421, 21381, 5708, 2257, 19105, 41908, 39823, 1546, 19425, 6627, 9574, 2257, 12565, 33453, 6627, 9574, 2257, 1543, 15789, 1541, 34627, 11336, 8802, 47421, 4438, 34912, 11, 18888, 8225, 23197, 4511, 9254, 1486, 115, 24726, 12467, 24726, 1546, 33952, 23700, 1543, 50909 ], "temperature": 0, "avg_logprob": -0.05687351715870392, "compression_ratio": 1.247011952191235, "no_speech_prob": 1.509710065294989e-11, "source": "whisper" }, { "id": 6, "seek": 9654, "start": 108.22, "end": 118.18, "text": "测试正式开始。跟上一代视觉实验版相比,在第二轮的测试里,中文写作和实图能力持平,前端视觉有一定的提升。", "tokens": [ 50949, 11038, 233, 5233, 243, 15789, 27584, 45213, 1543, 9678, 5708, 2257, 19105, 40656, 24447, 24726, 49657, 234, 42096, 15106, 11706, 11, 3581, 19693, 17819, 106, 1546, 11038, 233, 5233, 243, 15759, 11, 5975, 17174, 5676, 247, 11914, 12565, 24726, 3919, 122, 8225, 13486, 17694, 16716, 11, 8945, 11957, 107, 40656, 24447, 2412, 24272, 1546, 20949, 41670, 1543, 51447 ], "temperature": 0, "avg_logprob": -0.05687351715870392, "compression_ratio": 1.247011952191235, "no_speech_prob": 1.509710065294989e-11, "source": "whisper" }, { "id": 7, "seek": 11818, "start": 118.18, "end": 127.88000000000001, "text": "但我不得不收回它能达到前端Flash模型天花板的话。办公任务依然稳如老狗,而全战任务跟视觉实验版刚好打平。", "tokens": [ 50365, 8395, 25395, 5916, 1960, 18681, 8350, 11284, 8225, 9830, 122, 4511, 8945, 11957, 107, 25680, 1299, 41908, 39823, 6135, 20127, 43664, 44575, 1543, 39453, 13545, 26443, 5087, 94, 3254, 251, 5823, 10415, 111, 8238, 10439, 18637, 245, 11, 11070, 11319, 1486, 246, 26443, 5087, 94, 9678, 40656, 24447, 24726, 49657, 234, 42096, 49160, 2131, 12467, 16716, 1543, 50850 ], "temperature": 0, "avg_logprob": -0.04734768253741878, "compression_ratio": 1.2706185567010309, "no_speech_prob": 1.5634051345192468e-11, "source": "whisper" }, { "id": 8, "seek": 11818, "start": 128.46, "end": 135.36, "text": "具体的分数和计费我都列在屏幕上了。这些测试往期视频讲了太多次,今天只挑两个给大家看看。", "tokens": [ 50879, 39806, 29485, 1546, 6627, 33188, 12565, 7422, 94, 18464, 117, 1654, 7182, 43338, 3581, 9636, 237, 40211, 5708, 2289, 1543, 5562, 13824, 11038, 233, 5233, 243, 29510, 16786, 40656, 39752, 39255, 2289, 9455, 6392, 9487, 11, 12074, 14003, 49067, 36257, 7549, 23197, 6868, 22884, 1543, 51224 ], "temperature": 0, "avg_logprob": -0.04734768253741878, "compression_ratio": 1.2706185567010309, "no_speech_prob": 1.5634051345192468e-11, "source": "whisper" }, { "id": 9, "seek": 11818, "start": 135.98000000000002, "end": 139.38, "text": "先看前端能力,这次还是用外滩的昼夜风景来测试。", "tokens": [ 51255, 10108, 4200, 8945, 11957, 107, 8225, 13486, 11, 5562, 9487, 45726, 9254, 12022, 15868, 102, 1546, 1431, 120, 30124, 47209, 50218, 6912, 11038, 233, 5233, 243, 1543, 51425 ], "temperature": 0, "avg_logprob": -0.04734768253741878, "compression_ratio": 1.2706185567010309, "no_speech_prob": 1.5634051345192468e-11, "source": "whisper" }, { "id": 10, "seek": 11818, "start": 139.66, "end": 147.96, "text": "这个版本比起上一代儿童涂鸦式的呈现已经强了很多,至少能分辨出厨房三件套了,昼夜光线的变化也比较连贯自然。", "tokens": [ 51439, 15368, 42096, 8802, 11706, 9147, 5708, 2257, 19105, 22933, 11957, 98, 35681, 224, 165, 116, 99, 27584, 1546, 3606, 230, 20204, 49161, 5702, 118, 2289, 20778, 11, 20844, 15686, 8225, 6627, 9830, 101, 7781, 5014, 101, 38242, 10960, 20485, 1881, 245, 2289, 11, 1431, 120, 30124, 20690, 16853, 123, 1546, 2129, 246, 23756, 6404, 11706, 9830, 225, 3316, 252, 18464, 107, 9722, 5823, 1543, 51854 ], "temperature": 0, "avg_logprob": -0.04734768253741878, "compression_ratio": 1.2706185567010309, "no_speech_prob": 1.5634051345192468e-11, "source": "whisper" }, { "id": 11, "seek": 14818, "start": 148.42000000000002, "end": 151.92000000000002, "text": "但是,它仍然是非常简单的几何图案堆叠。", "tokens": [ 50377, 11189, 11, 11284, 1550, 235, 5823, 1541, 14392, 11249, 222, 47446, 1546, 6336, 254, 11906, 3919, 122, 28899, 10726, 228, 2129, 254, 1543, 50552 ], "temperature": 0, "avg_logprob": -0.10261242515162418, "compression_ratio": 1.2704918032786885, "no_speech_prob": 1.3510827873153275e-11, "source": "whisper" }, { "id": 12, "seek": 14818, "start": 152.38, "end": 156.38, "text": "另外有一点让我很惊讶,就是它消耗了比原来多一倍的token。", "tokens": [ 50575, 26202, 32241, 12579, 33650, 46257, 4799, 232, 7422, 114, 11, 5620, 11284, 28837, 4450, 245, 2289, 11706, 19683, 6912, 6392, 2257, 35477, 1546, 83, 8406, 1543, 50775 ], "temperature": 0, "avg_logprob": -0.10261242515162418, "compression_ratio": 1.2704918032786885, "no_speech_prob": 1.3510827873153275e-11, "source": "whisper" }, { "id": 13, "seek": 14818, "start": 157.12, "end": 160.64000000000001, "text": "为了让大家有个直观印象,我把这个测试项目里的冠军给大家看一下。", "tokens": [ 50812, 13992, 2289, 33650, 6868, 2412, 7549, 16186, 42350, 35825, 45007, 11, 1654, 16075, 15368, 11038, 233, 5233, 243, 10178, 117, 11386, 15759, 1546, 5676, 254, 5676, 249, 23197, 6868, 28324, 1543, 50988 ], "temperature": 0, "avg_logprob": -0.10261242515162418, "compression_ratio": 1.2704918032786885, "no_speech_prob": 1.3510827873153275e-11, "source": "whisper" }, { "id": 14, "seek": 14818, "start": 161.20000000000002, "end": 164.54000000000002, "text": "这么一笔Deepseek的Flash模型的前端能力,还得加油啊。", "tokens": [ 51016, 31889, 2257, 5437, 242, 11089, 595, 405, 916, 1546, 25680, 1299, 41908, 39823, 1546, 8945, 11957, 107, 8225, 13486, 11, 14852, 5916, 19310, 4905, 1543, 51183 ], "temperature": 0, "avg_logprob": -0.10261242515162418, "compression_ratio": 1.2704918032786885, "no_speech_prob": 1.3510827873153275e-11, "source": "whisper" }, { "id": 15, "seek": 14818, "start": 164.76000000000002, "end": 166.36, "text": "大家看看我的保留项目。", "tokens": [ 51194, 6868, 22884, 14200, 24302, 24456, 10178, 117, 11386, 1543, 51274 ], "temperature": 0, "avg_logprob": -0.10261242515162418, "compression_ratio": 1.2704918032786885, "no_speech_prob": 1.3510827873153275e-11, "source": "whisper" }, { "id": 16, "seek": 14818, "start": 166.8, "end": 175.34, "text": "这是一个复杂的长城任务,需要处理大约1G的原始素材,从零搭建数据库,构建整个小说阅读站的前后端代码。", "tokens": [ 51296, 27455, 20182, 1787, 235, 4422, 224, 1546, 32271, 45277, 26443, 5087, 94, 11, 35748, 1787, 226, 13876, 3582, 16853, 99, 16, 38, 1546, 19683, 15476, 37204, 4422, 238, 11, 35630, 6306, 114, 22220, 255, 34157, 33188, 26075, 106, 6346, 241, 11, 7360, 226, 34157, 27662, 7549, 7322, 8090, 10034, 227, 5233, 119, 34155, 1546, 8945, 13547, 11957, 107, 19105, 23230, 223, 1543, 51723 ], "temperature": 0, "avg_logprob": -0.10261242515162418, "compression_ratio": 1.2704918032786885, "no_speech_prob": 1.3510827873153275e-11, "source": "whisper" }, { "id": 17, "seek": 17534, "start": 175.34, "end": 181.74, "text": "4.1,Flash的速度极度亮眼,21分钟就完成了,燃烧了3210万token。", "tokens": [ 50365, 19, 13, 16, 11, 25680, 1299, 1546, 31217, 13127, 7360, 223, 13127, 22991, 25281, 11, 4436, 6627, 50064, 3111, 41509, 2289, 11, 24184, 225, 23661, 100, 2289, 11440, 3279, 23570, 83, 8406, 1543, 50685 ], "temperature": 0, "avg_logprob": -0.04476368272459352, "compression_ratio": 1.0, "no_speech_prob": 1.294258970829798e-11, "source": "whisper" }, { "id": 18, "seek": 17534, "start": 182.22, "end": 187.96, "text": "它的实现我仔细看了一遍,有三个明显的前端bug和一些其他小问题,大家可以看屏幕。", "tokens": [ 50709, 45224, 24726, 20204, 1654, 49385, 10115, 228, 4200, 2289, 2257, 3330, 235, 11, 2412, 10960, 7549, 11100, 1431, 122, 1546, 8945, 11957, 107, 44455, 12565, 38515, 9572, 5000, 7322, 34069, 11, 6868, 6723, 4200, 9636, 237, 40211, 1543, 50996 ], "temperature": 0, "avg_logprob": -0.04476368272459352, "compression_ratio": 1.0, "no_speech_prob": 1.294258970829798e-11, "source": "whisper" }, { "id": 19, "seek": 18796, "start": 187.96, "end": 193.56, "text": "这模型干活的速度可以说是快如闪电,但做事的脾气却非常中规中矩。", "tokens": [ 50365, 5562, 41908, 39823, 26111, 25956, 1546, 31217, 13127, 6723, 8090, 1541, 10251, 8238, 8259, 103, 42182, 11, 8395, 10907, 6973, 1546, 27067, 122, 42204, 5322, 112, 14392, 5975, 6758, 226, 5975, 5881, 102, 1543, 50645 ], "temperature": 0, "avg_logprob": -0.05042843740494525, "compression_ratio": 1.213235294117647, "no_speech_prob": 1.3049951744226185e-11, "source": "whisper" }, { "id": 20, "seek": 18796, "start": 194.04000000000002, "end": 198.16, "text": "它能很稳地实现你的要求,但绝不会多做任何需求之外的事情。", "tokens": [ 50669, 11284, 8225, 4563, 10415, 111, 10928, 24726, 20204, 18961, 4275, 32718, 11, 8395, 10115, 251, 47928, 6392, 10907, 26443, 11906, 32535, 32718, 9574, 12022, 1546, 24675, 1543, 50875 ], "temperature": 0, "avg_logprob": -0.05042843740494525, "compression_ratio": 1.213235294117647, "no_speech_prob": 1.3049951744226185e-11, "source": "whisper" }, { "id": 21, "seek": 18796, "start": 198.64000000000001, "end": 207.02, "text": "跟上一代视觉实验版相比,这个项目两边的分数完全持平,所以我是有点惊讶的,因为按照官方的说法,理应有很大的提升。", "tokens": [ 50899, 9678, 5708, 2257, 19105, 40656, 24447, 24726, 49657, 234, 42096, 15106, 11706, 11, 15368, 10178, 117, 11386, 36257, 40503, 1546, 6627, 33188, 37100, 17694, 16716, 11, 7239, 15914, 2412, 12579, 4799, 232, 7422, 114, 1546, 11, 34627, 26613, 32150, 31929, 9249, 1546, 8090, 11148, 11, 13876, 44297, 2412, 4563, 39156, 20949, 41670, 1543, 51318 ], "temperature": 0, "avg_logprob": -0.05042843740494525, "compression_ratio": 1.213235294117647, "no_speech_prob": 1.3049951744226185e-11, "source": "whisper" }, { "id": 22, "seek": 20702, "start": 207.02, "end": 212.0, "text": "来聊聊今天我想吐槽的第一个问题,4.1,Flash,真能代替Pro吗?", "tokens": [ 50365, 6912, 40096, 40096, 12074, 25246, 2392, 238, 24880, 121, 1546, 18049, 7549, 34069, 11, 19, 13, 16, 11, 25680, 1299, 11, 6303, 8225, 19105, 9531, 123, 12681, 14769, 30, 50614 ], "temperature": 0, "avg_logprob": -0.08168150583902994, "compression_ratio": 1.2294117647058824, "no_speech_prob": 1.3320232070124227e-11, "source": "whisper" }, { "id": 23, "seek": 20702, "start": 212.48000000000002, "end": 220.24, "text": "按照这个表格,Flash更擅长前端任务和Pro本身不支持的实图,而Pro在复杂任务上仍然稳稳地压过Flash。", "tokens": [ 50638, 26613, 32150, 15368, 17571, 30921, 11, 25680, 1299, 19002, 13167, 227, 32271, 8945, 11957, 107, 26443, 5087, 94, 12565, 12681, 8802, 19847, 1960, 35488, 1546, 24726, 3919, 122, 11, 11070, 12681, 3581, 1787, 235, 4422, 224, 26443, 5087, 94, 5708, 1550, 235, 5823, 10415, 111, 10415, 111, 10928, 5014, 233, 16866, 25680, 1299, 1543, 51026 ], "temperature": 0, "avg_logprob": -0.08168150583902994, "compression_ratio": 1.2294117647058824, "no_speech_prob": 1.3320232070124227e-11, "source": "whisper" }, { "id": 24, "seek": 20702, "start": 220.32000000000002, "end": 224.46, "text": "尤其是中文写作上的差距,直接让我痛失了最常用的写作模型。", "tokens": [ 51030, 1530, 97, 9572, 1541, 5975, 17174, 5676, 247, 11914, 5708, 1546, 21679, 6563, 251, 11, 43297, 33650, 1654, 28737, 28564, 2289, 8661, 11279, 9254, 1546, 5676, 247, 11914, 41908, 39823, 1543, 51237 ], "temperature": 0, "avg_logprob": -0.08168150583902994, "compression_ratio": 1.2294117647058824, "no_speech_prob": 1.3320232070124227e-11, "source": "whisper" }, { "id": 25, "seek": 20702, "start": 224.94, "end": 232.20000000000002, "text": "当然,对于每天用模型真正干活的人来说,速度也是非常重要的,等10分钟和等半小时,感受完全不一样。", "tokens": [ 51261, 40486, 11, 8713, 37732, 23664, 6135, 9254, 41908, 39823, 6303, 15789, 26111, 25956, 29979, 6912, 8090, 11, 31217, 13127, 22021, 14392, 24928, 1546, 11, 10187, 3279, 6627, 50064, 12565, 10187, 30018, 7322, 15729, 11, 9709, 23151, 37100, 1960, 2257, 14496, 1543, 51624 ], "temperature": 0, "avg_logprob": -0.08168150583902994, "compression_ratio": 1.2294117647058824, "no_speech_prob": 1.3320232070124227e-11, "source": "whisper" }, { "id": 26, "seek": 23220, "start": 232.2, "end": 241.67999999999998, "text": "所以在我看来,Pro和Flash其实是非常互补,缺一不可的,快速出前端视觉用Flash,而创意写作和复杂任务用Pro保证出品质量。", "tokens": [ 50365, 7239, 3581, 1654, 4200, 6912, 11, 12681, 12565, 25680, 1299, 41646, 1541, 14392, 1369, 240, 9890, 98, 11, 38109, 118, 2257, 1960, 4429, 1546, 11, 10251, 31217, 7781, 8945, 11957, 107, 40656, 24447, 9254, 25680, 1299, 11, 11070, 2437, 249, 9042, 5676, 247, 11914, 12565, 1787, 235, 4422, 224, 26443, 5087, 94, 9254, 12681, 24302, 5233, 223, 7781, 30246, 18464, 101, 26748, 1543, 50839 ], "temperature": 0, "avg_logprob": -0.07192181162751479, "compression_ratio": 1.2425068119891007, "no_speech_prob": 1.3165286569249979e-11, "source": "whisper" }, { "id": 27, "seek": 23220, "start": 242.29999999999998, "end": 244.57999999999998, "text": "那为什么Deepseek还要强行把Pro下架?", "tokens": [ 50870, 4184, 31006, 11089, 595, 405, 916, 14852, 4275, 5702, 118, 8082, 16075, 12681, 4438, 7360, 114, 30, 50984 ], "temperature": 0, "avg_logprob": -0.07192181162751479, "compression_ratio": 1.2425068119891007, "no_speech_prob": 1.3165286569249979e-11, "source": "whisper" }, { "id": 28, "seek": 23220, "start": 245.01999999999998, "end": 250.06, "text": "看到这个消息的时候,我真的是两眼一黑,而且并不是只有我一个人有这种体会。", "tokens": [ 51006, 18032, 15368, 28837, 26460, 49873, 11, 1654, 45698, 36257, 25281, 2257, 32141, 11, 22942, 3509, 114, 7296, 35244, 1654, 20182, 4035, 2412, 5562, 39810, 29485, 12949, 1543, 51258 ], "temperature": 0, "avg_logprob": -0.07192181162751479, "compression_ratio": 1.2425068119891007, "no_speech_prob": 1.3165286569249979e-11, "source": "whisper" }, { "id": 29, "seek": 23220, "start": 250.29999999999998, "end": 259.18, "text": "大模型社区里早就炒番了,少部分人是很开心的,因为他们很期待又快又有多模态的新款Flash,但另一派则是非常愤怒。", "tokens": [ 51270, 3582, 41908, 39823, 27658, 9937, 118, 15759, 21176, 3111, 4804, 240, 33118, 2289, 11, 15686, 32174, 4035, 1541, 4563, 18937, 7945, 1546, 11, 34627, 47911, 4563, 16786, 18390, 17047, 10251, 17047, 2412, 6392, 41908, 3757, 223, 1546, 12560, 48798, 25680, 1299, 11, 8395, 22762, 2257, 12136, 122, 2437, 247, 1541, 14392, 4820, 97, 3757, 240, 1543, 51714 ], "temperature": 0, "avg_logprob": -0.07192181162751479, "compression_ratio": 1.2425068119891007, "no_speech_prob": 1.3165286569249979e-11, "source": "whisper" }, { "id": 30, "seek": 25918, "start": 259.18, "end": 266.36, "text": "因为Pro模型的世界知识,更好的创意写作,以及在复杂任务里的稳定性,就这么被硬生生剥夺了。", "tokens": [ 50365, 34627, 12681, 41908, 39823, 1546, 24486, 6498, 5233, 228, 11, 19002, 20715, 2437, 249, 9042, 5676, 247, 11914, 11, 40282, 3581, 1787, 235, 4422, 224, 26443, 5087, 94, 15759, 1546, 10415, 111, 12088, 21686, 11, 3111, 31889, 23238, 38114, 105, 8244, 8244, 5935, 98, 1787, 118, 2289, 1543, 50724 ], "temperature": 0, "avg_logprob": -0.07212821719715896, "compression_ratio": 1.2275, "no_speech_prob": 1.496160660641177e-11, "source": "whisper" }, { "id": 31, "seek": 25918, "start": 266.72, "end": 269.0, "text": "而且这事只给了4天的时间做缓冲。", "tokens": [ 50742, 22942, 5562, 6973, 14003, 23197, 2289, 19, 6135, 1546, 44848, 10907, 38109, 241, 5676, 110, 1543, 50856 ], "temperature": 0, "avg_logprob": -0.07212821719715896, "compression_ratio": 1.2275, "no_speech_prob": 1.496160660641177e-11, "source": "whisper" }, { "id": 32, "seek": 25918, "start": 269.48, "end": 274.3, "text": "按照正常的商业逻辑,这种级别的切换,不应该至少给一个月的牵移期吗?", "tokens": [ 50880, 26613, 32150, 15789, 11279, 1546, 45581, 940, 248, 2215, 119, 9830, 239, 11, 5562, 39810, 16853, 100, 18453, 1546, 23632, 26075, 95, 11, 1960, 44297, 44646, 20844, 15686, 23197, 20182, 6939, 1546, 6935, 113, 8204, 119, 16786, 14769, 30, 51121 ], "temperature": 0, "avg_logprob": -0.07212821719715896, "compression_ratio": 1.2275, "no_speech_prob": 1.496160660641177e-11, "source": "whisper" }, { "id": 33, "seek": 25918, "start": 274.7, "end": 282.44, "text": "最后再多点我个人的观点,首先,恭喜Deepseek跑通了优秀的架构,首个多模态Flash上线,而且确实速度快效率高。", "tokens": [ 51141, 8661, 13547, 8623, 6392, 12579, 1654, 7549, 4035, 1546, 42350, 12579, 11, 36490, 11, 42179, 11089, 595, 405, 916, 32585, 19550, 2289, 7384, 246, 8204, 222, 1546, 7360, 114, 7360, 226, 11, 25444, 7549, 6392, 41908, 3757, 223, 25680, 1299, 5708, 16853, 123, 11, 22942, 38114, 106, 24726, 31217, 13127, 10251, 43076, 44866, 12979, 1543, 51528 ], "temperature": 0, "avg_logprob": -0.07212821719715896, "compression_ratio": 1.2275, "no_speech_prob": 1.496160660641177e-11, "source": "whisper" }, { "id": 34, "seek": 25918, "start": 282.88, "end": 287.32, "text": "但另一边,那只Pro大鲸鱼就这么被血藏了,这件事我心里确实过不去。", "tokens": [ 51550, 8395, 22762, 2257, 40503, 11, 4184, 14003, 12681, 3582, 165, 110, 116, 165, 109, 120, 3111, 31889, 23238, 44110, 21782, 237, 2289, 11, 5562, 20485, 6973, 1654, 7945, 15759, 38114, 106, 24726, 16866, 1960, 6734, 1543, 51772 ], "temperature": 0, "avg_logprob": -0.07212821719715896, "compression_ratio": 1.2275, "no_speech_prob": 1.496160660641177e-11, "source": "whisper" }, { "id": 35, "seek": 28732, "start": 287.32, "end": 296.84, "text": "作为一个每天都在跟大模型打交道的程序员,我已经把自动脚本里的Pro模型迁移到了其他开源模型供应商,但仍然觉得这个切换太过于仓促了。", "tokens": [ 50365, 11914, 13992, 20182, 23664, 6135, 7182, 3581, 9678, 3582, 41908, 39823, 12467, 28455, 6025, 1546, 29649, 6346, 237, 3606, 246, 11, 1654, 49161, 16075, 9722, 34961, 27067, 248, 8802, 15759, 1546, 12681, 41908, 39823, 3316, 223, 8204, 119, 21381, 9572, 5000, 18937, 47402, 41908, 39823, 3254, 249, 44297, 45581, 11, 8395, 1550, 235, 5823, 29992, 15368, 23632, 26075, 95, 9455, 16866, 37732, 1550, 241, 7792, 225, 2289, 1543, 50841 ], "temperature": 0, "avg_logprob": -0.06274898891342419, "compression_ratio": 1.2440318302387268, "no_speech_prob": 1.3993825195846021e-11, "source": "whisper" }, { "id": 36, "seek": 28732, "start": 297.34, "end": 302.4, "text": "另外呢,平心而论,只靠Flash模型是撑不起一家大模型公司的未来的。", "tokens": [ 50866, 26202, 6240, 11, 16716, 7945, 11070, 7422, 118, 11, 14003, 5363, 254, 25680, 1299, 41908, 39823, 1541, 20559, 239, 26451, 2257, 5155, 3582, 41908, 39823, 13545, 32981, 1546, 29954, 6912, 1546, 1543, 51119 ], "temperature": 0, "avg_logprob": -0.06274898891342419, "compression_ratio": 1.2440318302387268, "no_speech_prob": 1.3993825195846021e-11, "source": "whisper" }, { "id": 37, "seek": 28732, "start": 302.71999999999997, "end": 308.21999999999997, "text": "所以,我个人非常非常希望能看到公告里说的4.1Pro尽快上线。", "tokens": [ 51135, 7239, 11, 1654, 7549, 4035, 14392, 14392, 29955, 8225, 18032, 13545, 16846, 15759, 8090, 1546, 19, 13, 16, 12681, 1530, 121, 10251, 5708, 16853, 123, 1543, 51410 ], "temperature": 0, "avg_logprob": -0.06274898891342419, "compression_ratio": 1.2440318302387268, "no_speech_prob": 1.3993825195846021e-11, "source": "whisper" }, { "id": 38, "seek": 28732, "start": 308.5, "end": 314.56, "text": "如果我说的不对,欢迎在评论区指证,每次我都能在那儿看到特别精彩的发言,就等你来哦。", "tokens": [ 51424, 13119, 1654, 8090, 1546, 41639, 11, 28566, 17699, 3581, 5233, 226, 7422, 118, 9937, 118, 25922, 5233, 223, 11, 23664, 9487, 1654, 7182, 8225, 3581, 4184, 22933, 18032, 17682, 18453, 34910, 7391, 102, 1546, 28926, 12009, 11, 3111, 10187, 2166, 6912, 14659, 1543, 51727 ], "temperature": 0, "avg_logprob": -0.06274898891342419, "compression_ratio": 1.2440318302387268, "no_speech_prob": 1.3993825195846021e-11, "source": "whisper" } ], "transcript_selection": { "schema": "videonote.transcript_selection.v1", "status": "SELECTED", "requested_language": "zh", "duration": 314.93225, "reference_path": null, "reference_kind": null, "reference_language": null, "selected_source": "whisper_segmented", "decision_reason": "source_unavailable_or_untrusted_whisper_passed", "whisper_rejected": false, "source_profile": { "segment_count": 0, "duration_coverage": 0.0, "usable_full": false, "usable_partial": false, "hallucination_region_count": 0, "hallucination_regions": [], "language": { "requested": "zh", "matched": false } }, "whisper_profile": { "segment_count": 39, "multilingual": false, "first_start": 0.0, "final_end": 314.56, "duration_coverage": 0.998818, "maximum_gap_seconds": 0.8, "inverted_count": 0, "language": { "requested": "zh", "matched": true, "latin": 209, "han": 1531, "kana": 0, "hangul": 0 }, "hallucination_region_count": 0, "hallucination_regions": [], "usable_full": true, "usable_partial": true }, "comparison": { "window_seconds": 30.0, "compared_window_count": 0, "median_score": 0.0, "divergent_window_count": 0, "divergent_ratio": 0.0, "divergent_windows": [] }, "timestamp_alignment": { "applied": false, "estimated_offset_seconds": 0.0, "matched_segment_count": 0, "match_scores": [] }, "caption_timing_repair": null, "filled_source_profile": null, "whisper_duration_guard": { "media_duration": 314.932, "rejected_segment_count": 21, "rejected_segments": [ { "index": 39, "start": 314.56, "end": 316.56, "reason": "material_end_timestamp_overrun" }, { "index": 40, "start": 316.56, "end": 317.56, "reason": "starts_at_or_after_media_end" }, { "index": 41, "start": 317.56, "end": 318.56, "reason": "starts_at_or_after_media_end" }, { "index": 42, "start": 318.56, "end": 319.56, "reason": "starts_at_or_after_media_end" }, { "index": 43, "start": 319.56, "end": 320.56, "reason": "starts_at_or_after_media_end" }, { "index": 44, "start": 320.56, "end": 321.56, "reason": "starts_at_or_after_media_end" }, { "index": 45, "start": 321.56, "end": 322.56, "reason": "starts_at_or_after_media_end" }, { "index": 46, "start": 322.56, "end": 324.56, "reason": "starts_at_or_after_media_end" }, { "index": 47, "start": 324.56, "end": 325.56, "reason": "starts_at_or_after_media_end" }, { "index": 48, "start": 325.56, "end": 326.56, "reason": "starts_at_or_after_media_end" }, { "index": 49, "start": 326.56, "end": 327.56, "reason": "starts_at_or_after_media_end" }, { "index": 50, "start": 327.56, "end": 328.56, "reason": "starts_at_or_after_media_end" }, { "index": 51, "start": 328.56, "end": 329.56, "reason": "starts_at_or_after_media_end" }, { "index": 52, "start": 329.56, "end": 330.56, "reason": "starts_at_or_after_media_end" }, { "index": 53, "start": 330.56, "end": 331.56, "reason": "starts_at_or_after_media_end" }, { "index": 54, "start": 331.56, "end": 332.56, "reason": "starts_at_or_after_media_end" }, { "index": 55, "start": 332.56, "end": 333.56, "reason": "starts_at_or_after_media_end" }, { "index": 56, "start": 333.56, "end": 334.56, "reason": "starts_at_or_after_media_end" }, { "index": 57, "start": 334.56, "end": 335.56, "reason": "starts_at_or_after_media_end" }, { "index": 58, "start": 335.56, "end": 336.56, "reason": "starts_at_or_after_media_end" }, { "index": 59, "start": 336.56, "end": 337.56, "reason": "starts_at_or_after_media_end" } ] } } }