{"text": "视频还没播完,MOSVIA Real Time已经开始描述画面。他先认出咖啡店和走进门的一男一女,随后又跟到了柜台前的交谈。后面的人物继续出现,文字也继续往下涨。这不是等视频结束以后再做总结,模型正在跟着画面一起走。它是MOS智能推出的110亿参数开放全入模型,专门处理持续进入的视频流。普通的视频问答往往要先看完整段再统一回答。事实版把等待拆开了,新画面不断进来,理解和文字生成也可以同时继续。按照官方介绍,你可以在视频播放过程中插入问题、证据还不够时。它可以先保持沉默,后面的画面改变了情境,它还能补充或者修正此前的判断。边看边答,主动沉默和动态更新构成了这个版本最核心的区别。同系列还有基础版和指令版。基础版适合继续训练,指令版面向完整视频的深入问答。实施版才专门面对持续视频流。官方标注的256000上下文,表示一次能够参考的信息容量,不是固定的视频分钟数。抽针越密,对话越长,能够容纳的视频时长就会跟着变化。接下来是我最关心的问题,消费机显卡能不能把它跑起来。官方给出了4比特量化版本,按照一张R4G显卡设计,我没有这张卡,于是改用两张4070Ti Super让程序自动拆分模型。本地部署走的是一条完整的成功路径。仓库拉到本地以后,我创建独立环境,装好运行依赖,再下载官方量化权重。启动实施推理程序时,三个权重分片依次加载。加载完成以后,模型开始按时间接收视频。两张显卡也同时参与推理。画面继续进入,文字随之更新,终端最后显示36帧全部送达,这次本地部署正式跑通。测试输入是一段30多秒的咖啡店视频,程序每秒取一帧,模型先认出了门店和进场人物,随后描述男子走向柜台,店员递出纸杯,场景继续变化,他又跟到了后面的人物冲突。整段输出不是最后一次性出现,而是随着视频推进不断补充。我又在视频播到约10秒时插入一个问题,问现在画面里有几个人,他先回答两个人。7秒以后,柜台后的电源进入画面,他没有等我再问,直接把答案更新成3个人,到了26秒我再问刚才发生了什么变化。他先认出电源正在介绍三种杯子,随后又补充男子已经走到柜台前,两次提问都发生在播放过程中,后面的画面也没有被第一轮回答挡住。这次运行也给出了一个很直接的硬件结论,我用两张4070Ti Super跑完了这段视频,扣掉系统本来的占用,模型额外使用的显存大约13.3G,这个结果只对应当前样片和参数,但至少证明这套双卡配置可以完成一次本地实施推理。输出中间还有一段很值得看画面没有明显变化时,终端没有继续增加描述,后面的人物进入以后文字才重新出现。原始接口会把这种状态写成控制标记,真正做成应用时需要先过滤,对观众来说更重要的是模型知道什么时候先观察,而不是每一帧都重复说话。不过,实时不代表每个瞬间都不会漏掉,当前测试每秒只取一帧,视频里一个非常短的击打动作就没有被稳定识别。它看到了动作前后不等于看见了中间的每一刻,提高采样频率可能保留更多细节,也会同时增加输入和缓存压力。这种持续更新来自模型的输入方式,视觉信息不是等到整段结束再交给语言部分,每个采样画面还带着自己的时间位置,所以事件先后和状态变化能够进入同一段推理。官方代码除了读取本地视频还支持摄像头、屏幕和外部发送的视频帧,它可以成为直播辅助实验观察、流程提醒或者屏幕任务监控的视觉底座。模型负责持续理解和回答,真正上线是告紧规则、权限控制和人工复合仍然要由应用层补上。所以Mousewell Real-Time的优势很明确,视频还在发声,问题和答案就能一起向前推进。它的限制也同样明确,采样会漏掉短动作,长上下文会继续消耗缓存,原始输出也不是装好就能使用的监控产品。如果你的任务只是等视频结束以后再总结,指令版会更直接。如果你需要在视频进行中观察、提问等待,再根据新画面更新判断,实施版才真正对口。项目地址和官方资料放在评论区,感兴趣可以拿自己的视频试一遍。", "segments": [{"id": 0, "seek": 0, "start": 0.0, "end": 3.12, "text": "视频还没播完,MOSVIA Real Time已经开始描述画面。", "tokens": [50365, 40656, 39752, 14852, 10062, 49993, 14128, 11, 44, 4367, 53, 6914, 8467, 6161, 49161, 45213, 11673, 237, 3316, 108, 27126, 8833, 1543, 50521], "temperature": 0, "avg_logprob": -0.13174379252951984, "compression_ratio": 1.2915662650602409, "no_speech_prob": 7.634058316052794e-12}, {"id": 1, "seek": 0, "start": 3.12, "end": 7.8, "text": "他先认出咖啡店和走进门的一男一女,随后又跟到了柜台前的交谈。", "tokens": [50521, 5000, 10108, 7422, 97, 7781, 8975, 244, 3284, 94, 22469, 12565, 9575, 36700, 8259, 101, 1546, 2257, 25243, 2257, 17015, 11, 10673, 237, 13547, 17047, 9678, 21381, 17238, 250, 15433, 8945, 1546, 28455, 8897, 230, 1543, 50755], "temperature": 0, "avg_logprob": -0.13174379252951984, "compression_ratio": 1.2915662650602409, "no_speech_prob": 7.634058316052794e-12}, {"id": 2, "seek": 0, "start": 7.8, "end": 11.200000000000001, "text": "后面的人物继续出现,文字也继续往下涨。", "tokens": [50755, 13547, 8833, 29979, 23516, 10115, 100, 10115, 255, 7781, 20204, 11, 17174, 22381, 6404, 10115, 100, 10115, 255, 29510, 4438, 35681, 101, 1543, 50925], "temperature": 0, "avg_logprob": -0.13174379252951984, "compression_ratio": 1.2915662650602409, "no_speech_prob": 7.634058316052794e-12}, {"id": 3, "seek": 0, "start": 11.200000000000001, "end": 15.52, "text": "这不是等视频结束以后再做总结,模型正在跟着画面一起走。", "tokens": [50925, 5562, 7296, 10187, 40656, 39752, 45641, 44320, 3588, 13547, 8623, 10907, 33440, 45641, 11, 41908, 39823, 15789, 3581, 9678, 20708, 27126, 8833, 29567, 9575, 1543, 51141], "temperature": 0, "avg_logprob": -0.13174379252951984, "compression_ratio": 1.2915662650602409, "no_speech_prob": 7.634058316052794e-12}, {"id": 4, "seek": 0, "start": 15.52, "end": 21.2, "text": "它是MOS智能推出的110亿参数开放全入模型,专门处理持续进入的视频流。", "tokens": [51141, 11284, 1541, 44, 4367, 5094, 118, 8225, 33597, 7781, 1546, 43435, 1369, 123, 2129, 224, 33188, 18937, 12744, 11319, 14028, 41908, 39823, 11, 940, 241, 8259, 101, 1787, 226, 13876, 17694, 10115, 255, 36700, 14028, 1546, 40656, 39752, 27854, 1543, 51425], "temperature": 0, "avg_logprob": -0.13174379252951984, "compression_ratio": 1.2915662650602409, "no_speech_prob": 7.634058316052794e-12}, {"id": 5, "seek": 0, "start": 21.2, "end": 24.6, "text": "普通的视频问答往往要先看完整段再统一回答。", "tokens": [51425, 29993, 19550, 1546, 40656, 39752, 22064, 28223, 29510, 29510, 4275, 10108, 4200, 14128, 27662, 28427, 8623, 10115, 253, 2257, 8350, 28223, 1543, 51595], "temperature": 0, "avg_logprob": -0.13174379252951984, "compression_ratio": 1.2915662650602409, "no_speech_prob": 7.634058316052794e-12}, {"id": 6, "seek": 0, "start": 24.6, "end": 29.68, "text": "事实版把等待拆开了,新画面不断进来,理解和文字生成也可以同时继续。", "tokens": [51595, 6973, 24726, 42096, 16075, 10187, 18390, 6852, 228, 18937, 2289, 11, 12560, 27126, 8833, 1960, 4307, 255, 36700, 6912, 11, 13876, 17278, 12565, 17174, 22381, 8244, 11336, 47935, 13089, 15729, 10115, 100, 10115, 255, 1543, 51849], "temperature": 0, "avg_logprob": -0.13174379252951984, "compression_ratio": 1.2915662650602409, "no_speech_prob": 7.634058316052794e-12}, {"id": 7, "seek": 2968, "start": 29.68, "end": 34.68, "text": "按照官方介绍,你可以在视频播放过程中插入问题、证据还不够时。", "tokens": [50365, 26613, 32150, 31929, 9249, 30312, 10115, 235, 11, 42766, 3581, 40656, 39752, 49993, 12744, 16866, 29649, 5975, 11673, 240, 14028, 34069, 1231, 5233, 223, 26075, 106, 14852, 1960, 1787, 253, 15729, 1543, 50615], "temperature": 0, "avg_logprob": -0.07606139341236856, "compression_ratio": 1.3478260869565217, "no_speech_prob": 1.428204429720914e-11}, {"id": 8, "seek": 2968, "start": 34.68, "end": 39.88, "text": "它可以先保持沉默,后面的画面改变了情境,它还能补充或者修正此前的判断。", "tokens": [50615, 11284, 6723, 10108, 24302, 17694, 3308, 231, 6173, 246, 11, 13547, 8833, 1546, 27126, 8833, 34490, 2129, 246, 2289, 10570, 47885, 11, 11284, 14852, 8225, 9890, 98, 2347, 227, 31148, 7792, 106, 15789, 17947, 8945, 1546, 2437, 97, 4307, 255, 1543, 50875], "temperature": 0, "avg_logprob": -0.07606139341236856, "compression_ratio": 1.3478260869565217, "no_speech_prob": 1.428204429720914e-11}, {"id": 9, "seek": 2968, "start": 39.88, "end": 44.28, "text": "边看边答,主动沉默和动态更新构成了这个版本最核心的区别。", "tokens": [50875, 40503, 4200, 40503, 28223, 11, 13557, 34961, 3308, 231, 6173, 246, 12565, 34961, 3757, 223, 19002, 12560, 7360, 226, 11336, 2289, 15368, 42096, 8802, 8661, 8987, 116, 7945, 1546, 9937, 118, 18453, 1543, 51095], "temperature": 0, "avg_logprob": -0.07606139341236856, "compression_ratio": 1.3478260869565217, "no_speech_prob": 1.428204429720914e-11}, {"id": 10, "seek": 2968, "start": 44.28, "end": 46.480000000000004, "text": "同系列还有基础版和指令版。", "tokens": [51095, 13089, 25368, 43338, 35091, 26008, 38114, 222, 42096, 12565, 25922, 49061, 42096, 1543, 51205], "temperature": 0, "avg_logprob": -0.07606139341236856, "compression_ratio": 1.3478260869565217, "no_speech_prob": 1.428204429720914e-11}, {"id": 11, "seek": 2968, "start": 46.480000000000004, "end": 50.6, "text": "基础版适合继续训练,指令版面向完整视频的深入问答。", "tokens": [51205, 26008, 38114, 222, 42096, 2215, 224, 14245, 10115, 100, 10115, 255, 7422, 255, 10115, 225, 11, 25922, 49061, 42096, 8833, 24282, 14128, 27662, 40656, 39752, 1546, 24043, 14028, 22064, 28223, 1543, 51411], "temperature": 0, "avg_logprob": -0.07606139341236856, "compression_ratio": 1.3478260869565217, "no_speech_prob": 1.428204429720914e-11}, {"id": 12, "seek": 2968, "start": 50.6, "end": 53.2, "text": "实施版才专门面对持续视频流。", "tokens": [51411, 24726, 4307, 121, 42096, 18888, 940, 241, 8259, 101, 8833, 8713, 17694, 10115, 255, 40656, 39752, 27854, 1543, 51541], "temperature": 0, "avg_logprob": -0.07606139341236856, "compression_ratio": 1.3478260869565217, "no_speech_prob": 1.428204429720914e-11}, {"id": 13, "seek": 2968, "start": 53.2, "end": 58.08, "text": "官方标注的256000上下文,表示一次能够参考的信息容量,", "tokens": [51541, 31929, 9249, 162, 3921, 26432, 1546, 6074, 21, 1360, 5708, 4438, 17174, 11, 40053, 27505, 8225, 1787, 253, 2129, 224, 26504, 1546, 17665, 26460, 25750, 26748, 11, 51785], "temperature": 0, "avg_logprob": -0.07606139341236856, "compression_ratio": 1.3478260869565217, "no_speech_prob": 1.428204429720914e-11}, {"id": 14, "seek": 5808, "start": 58.08, "end": 64.67999999999999, "text": "不是固定的视频分钟数。抽针越密,对话越长,能够容纳的视频时长就会跟着变化。", "tokens": [50365, 7296, 3919, 118, 12088, 1546, 40656, 39752, 6627, 50064, 33188, 1543, 46022, 25153, 230, 25761, 47838, 11, 8713, 21596, 25761, 32271, 11, 8225, 1787, 253, 25750, 16853, 111, 1546, 40656, 39752, 15729, 32271, 3111, 12949, 9678, 20708, 2129, 246, 23756, 1543, 50695], "temperature": 0, "avg_logprob": -0.13445081443430107, "compression_ratio": 1.162280701754386, "no_speech_prob": 1.4966614753086915e-11}, {"id": 15, "seek": 5808, "start": 64.67999999999999, "end": 68.6, "text": "接下来是我最关心的问题,消费机显卡能不能把它跑起来。", "tokens": [50695, 14468, 4438, 6912, 42614, 8661, 28053, 7945, 1546, 34069, 11, 28837, 18464, 117, 37960, 1431, 122, 32681, 8225, 28590, 42061, 32585, 9147, 6912, 1543, 50891], "temperature": 0, "avg_logprob": -0.13445081443430107, "compression_ratio": 1.162280701754386, "no_speech_prob": 1.4966614753086915e-11}, {"id": 16, "seek": 5808, "start": 68.6, "end": 73.56, "text": "官方给出了4比特量化版本,按照一张R4G显卡设计,我没有这张卡,", "tokens": [50891, 31929, 9249, 23197, 7781, 2289, 19, 11706, 17682, 26748, 23756, 42096, 8802, 11, 26613, 32150, 2257, 44059, 49, 19, 38, 1431, 122, 32681, 7422, 122, 7422, 94, 11, 1654, 17944, 5562, 44059, 32681, 11, 51139], "temperature": 0, "avg_logprob": -0.13445081443430107, "compression_ratio": 1.162280701754386, "no_speech_prob": 1.4966614753086915e-11}, {"id": 17, "seek": 7356, "start": 73.56, "end": 78.16, "text": "于是改用两张4070Ti Super让程序自动拆分模型。", "tokens": [50365, 37732, 1541, 34490, 9254, 36257, 44059, 5254, 5867, 51, 72, 4548, 33650, 29649, 6346, 237, 9722, 34961, 6852, 228, 6627, 41908, 39823, 1543, 50595], "temperature": 0, "avg_logprob": -0.09501959572375661, "compression_ratio": 1.1702898550724639, "no_speech_prob": 1.4908263859525483e-11}, {"id": 18, "seek": 7356, "start": 78.16, "end": 81.16, "text": "本地部署走的是一条完整的成功路径。", "tokens": [50595, 8802, 10928, 13470, 16469, 110, 9575, 24620, 2257, 48837, 14128, 27662, 1546, 11336, 32311, 24658, 2172, 226, 1543, 50745], "temperature": 0, "avg_logprob": -0.09501959572375661, "compression_ratio": 1.1702898550724639, "no_speech_prob": 1.4908263859525483e-11}, {"id": 19, "seek": 7356, "start": 81.16, "end": 85.88, "text": "仓库拉到本地以后,我创建独立环境,装好运行依赖,再下载官方量化权重。", "tokens": [50745, 1550, 241, 6346, 241, 29717, 4511, 8802, 10928, 3588, 13547, 11, 1654, 2437, 249, 34157, 18637, 105, 24409, 8051, 107, 47885, 11, 8083, 227, 2131, 3316, 238, 8082, 3254, 251, 5266, 244, 11, 8623, 4438, 17819, 121, 31929, 9249, 26748, 23756, 4422, 225, 12624, 1543, 50981], "temperature": 0, "avg_logprob": -0.09501959572375661, "compression_ratio": 1.1702898550724639, "no_speech_prob": 1.4908263859525483e-11}, {"id": 20, "seek": 7356, "start": 85.88, "end": 89.04, "text": "启动实施推理程序时,三个权重分片依次加载。", "tokens": [50981, 2392, 107, 34961, 24726, 4307, 121, 33597, 13876, 29649, 6346, 237, 15729, 11, 10960, 7549, 4422, 225, 12624, 6627, 16668, 3254, 251, 9487, 9990, 17819, 121, 1543, 51139], "temperature": 0, "avg_logprob": -0.09501959572375661, "compression_ratio": 1.1702898550724639, "no_speech_prob": 1.4908263859525483e-11}, {"id": 21, "seek": 7356, "start": 89.04, "end": 92.16, "text": "加载完成以后,模型开始按时间接收视频。", "tokens": [51139, 9990, 17819, 121, 41509, 3588, 13547, 11, 41908, 39823, 45213, 26613, 44848, 14468, 18681, 40656, 39752, 1543, 51295], "temperature": 0, "avg_logprob": -0.09501959572375661, "compression_ratio": 1.1702898550724639, "no_speech_prob": 1.4908263859525483e-11}, {"id": 22, "seek": 9216, "start": 92.16, "end": 94.44, "text": "两张显卡也同时参与推理。", "tokens": [50365, 36257, 44059, 1431, 122, 32681, 6404, 13089, 15729, 2129, 224, 940, 236, 33597, 13876, 1543, 50479], "temperature": 0, "avg_logprob": -0.10659976996997796, "compression_ratio": 1.301775147928994, "no_speech_prob": 1.3896770019450333e-11}, {"id": 23, "seek": 9216, "start": 94.44, "end": 99.84, "text": "画面继续进入,文字随之更新,终端最后显示36帧全部送达,", "tokens": [50479, 27126, 8833, 10115, 100, 10115, 255, 36700, 14028, 11, 17174, 22381, 10673, 237, 9574, 19002, 12560, 11, 10115, 230, 11957, 107, 8661, 13547, 1431, 122, 25696, 11309, 4845, 100, 38714, 29309, 9830, 122, 11, 50749], "temperature": 0, "avg_logprob": -0.10659976996997796, "compression_ratio": 1.301775147928994, "no_speech_prob": 1.3896770019450333e-11}, {"id": 24, "seek": 9216, "start": 99.84, "end": 101.75999999999999, "text": "这次本地部署正式跑通。", "tokens": [50749, 5562, 9487, 8802, 10928, 13470, 16469, 110, 15789, 27584, 32585, 19550, 1543, 50845], "temperature": 0, "avg_logprob": -0.10659976996997796, "compression_ratio": 1.301775147928994, "no_speech_prob": 1.3896770019450333e-11}, {"id": 25, "seek": 9216, "start": 101.75999999999999, "end": 106.08, "text": "测试输入是一段30多秒的咖啡店视频,程序每秒取一帧,", "tokens": [50845, 11038, 233, 5233, 243, 9830, 241, 14028, 1541, 2257, 28427, 3446, 6392, 20956, 1546, 8975, 244, 3284, 94, 22469, 40656, 39752, 11, 29649, 6346, 237, 23664, 20956, 29436, 2257, 4845, 100, 11, 51061], "temperature": 0, "avg_logprob": -0.10659976996997796, "compression_ratio": 1.301775147928994, "no_speech_prob": 1.3896770019450333e-11}, {"id": 26, "seek": 9216, "start": 106.08, "end": 110.24, "text": "模型先认出了门店和进场人物,随后描述男子走向柜台,", "tokens": [51061, 41908, 39823, 10108, 7422, 97, 7781, 2289, 8259, 101, 22469, 12565, 36700, 50255, 4035, 23516, 11, 10673, 237, 13547, 11673, 237, 3316, 108, 25243, 7626, 9575, 24282, 17238, 250, 15433, 11, 51269], "temperature": 0, "avg_logprob": -0.10659976996997796, "compression_ratio": 1.301775147928994, "no_speech_prob": 1.3896770019450333e-11}, {"id": 27, "seek": 9216, "start": 110.24, "end": 115.24, "text": "店员递出纸杯,场景继续变化,他又跟到了后面的人物冲突。", "tokens": [51269, 22469, 3606, 246, 2215, 240, 7781, 16853, 116, 45458, 11, 50255, 50218, 10115, 100, 10115, 255, 2129, 246, 23756, 11, 5000, 17047, 9678, 21381, 13547, 8833, 29979, 23516, 5676, 110, 40859, 1543, 51519], "temperature": 0, "avg_logprob": -0.10659976996997796, "compression_ratio": 1.301775147928994, "no_speech_prob": 1.3896770019450333e-11}, {"id": 28, "seek": 9216, "start": 115.24, "end": 120.03999999999999, "text": "整段输出不是最后一次性出现,而是随着视频推进不断补充。", "tokens": [51519, 27662, 28427, 9830, 241, 7781, 7296, 8661, 13547, 27505, 21686, 7781, 20204, 11, 11070, 1541, 10673, 237, 20708, 40656, 39752, 33597, 36700, 1960, 4307, 255, 9890, 98, 2347, 227, 1543, 51759], "temperature": 0, "avg_logprob": -0.10659976996997796, "compression_ratio": 1.301775147928994, "no_speech_prob": 1.3896770019450333e-11}, {"id": 29, "seek": 12004, "start": 120.04, "end": 125.08000000000001, "text": "我又在视频播到约10秒时插入一个问题,问现在画面里有几个人,", "tokens": [50365, 1654, 17047, 3581, 40656, 39752, 49993, 4511, 16853, 99, 3279, 20956, 15729, 11673, 240, 14028, 20182, 34069, 11, 22064, 25040, 27126, 8833, 15759, 2412, 6336, 254, 7549, 4035, 11, 50617], "temperature": 0, "avg_logprob": -0.10400979895340769, "compression_ratio": 1.378917378917379, "no_speech_prob": 1.814225505936129e-11}, {"id": 30, "seek": 12004, "start": 125.08000000000001, "end": 126.24000000000001, "text": "他先回答两个人。", "tokens": [50617, 5000, 10108, 8350, 28223, 36257, 7549, 4035, 1543, 50675], "temperature": 0, "avg_logprob": -0.10400979895340769, "compression_ratio": 1.378917378917379, "no_speech_prob": 1.814225505936129e-11}, {"id": 31, "seek": 12004, "start": 126.24000000000001, "end": 129.84, "text": "7秒以后,柜台后的电源进入画面,他没有等我再问,", "tokens": [50675, 22, 20956, 3588, 13547, 11, 17238, 250, 15433, 13547, 1546, 42182, 47402, 36700, 14028, 27126, 8833, 11, 5000, 17944, 10187, 1654, 8623, 22064, 11, 50855], "temperature": 0, "avg_logprob": -0.10400979895340769, "compression_ratio": 1.378917378917379, "no_speech_prob": 1.814225505936129e-11}, {"id": 32, "seek": 12004, "start": 129.84, "end": 134.32, "text": "直接把答案更新成3个人,到了26秒我再问刚才发生了什么变化。", "tokens": [50855, 43297, 16075, 28223, 28899, 19002, 12560, 11336, 18, 7549, 4035, 11, 21381, 10880, 20956, 1654, 8623, 22064, 49160, 18888, 28926, 8244, 2289, 10440, 2129, 246, 23756, 1543, 51079], "temperature": 0, "avg_logprob": -0.10400979895340769, "compression_ratio": 1.378917378917379, "no_speech_prob": 1.814225505936129e-11}, {"id": 33, "seek": 12004, "start": 134.32, "end": 139.04000000000002, "text": "他先认出电源正在介绍三种杯子,随后又补充男子已经走到柜台前,", "tokens": [51079, 5000, 10108, 7422, 97, 7781, 42182, 47402, 15789, 3581, 30312, 10115, 235, 10960, 39810, 45458, 7626, 11, 10673, 237, 13547, 17047, 9890, 98, 2347, 227, 25243, 7626, 49161, 9575, 4511, 17238, 250, 15433, 8945, 11, 51315], "temperature": 0, "avg_logprob": -0.10400979895340769, "compression_ratio": 1.378917378917379, "no_speech_prob": 1.814225505936129e-11}, {"id": 34, "seek": 12004, "start": 139.04000000000002, "end": 143.72, "text": "两次提问都发生在播放过程中,后面的画面也没有被第一轮回答挡住。", "tokens": [51315, 36257, 9487, 20949, 22064, 7182, 28926, 8244, 3581, 49993, 12744, 16866, 29649, 5975, 11, 13547, 8833, 1546, 27126, 8833, 6404, 17944, 23238, 18049, 17819, 106, 8350, 28223, 8501, 94, 21632, 1543, 51549], "temperature": 0, "avg_logprob": -0.10400979895340769, "compression_ratio": 1.378917378917379, "no_speech_prob": 1.814225505936129e-11}, {"id": 35, "seek": 12004, "start": 143.72, "end": 147.16, "text": "这次运行也给出了一个很直接的硬件结论,", "tokens": [51549, 5562, 9487, 3316, 238, 8082, 6404, 23197, 7781, 2289, 20182, 4563, 43297, 1546, 38114, 105, 20485, 45641, 7422, 118, 11, 51721], "temperature": 0, "avg_logprob": -0.10400979895340769, "compression_ratio": 1.378917378917379, "no_speech_prob": 1.814225505936129e-11}, {"id": 36, "seek": 14716, "start": 147.16, "end": 152.28, "text": "我用两张4070Ti Super跑完了这段视频,扣掉系统本来的占用,", "tokens": [50365, 1654, 9254, 36257, 44059, 5254, 5867, 51, 72, 4548, 32585, 41665, 5562, 28427, 40656, 39752, 11, 3416, 96, 29327, 25368, 10115, 253, 8802, 6912, 1546, 5322, 254, 9254, 11, 50621], "temperature": 0, "avg_logprob": -0.08952753360454853, "compression_ratio": 1.147239263803681, "no_speech_prob": 1.6501086955167388e-11}, {"id": 37, "seek": 14716, "start": 152.28, "end": 157.84, "text": "模型额外使用的显存大约13.3G,这个结果只对应当前样片和参数,", "tokens": [50621, 41908, 39823, 12501, 251, 12022, 22982, 9254, 1546, 1431, 122, 39781, 3582, 16853, 99, 7668, 13, 18, 38, 11, 15368, 45641, 9319, 14003, 8713, 44297, 16233, 8945, 14496, 16668, 12565, 2129, 224, 33188, 11, 50899], "temperature": 0, "avg_logprob": -0.08952753360454853, "compression_ratio": 1.147239263803681, "no_speech_prob": 1.6501086955167388e-11}, {"id": 38, "seek": 14716, "start": 157.84, "end": 162.04, "text": "但至少证明这套双卡配置可以完成一次本地实施推理。", "tokens": [50899, 8395, 20844, 15686, 5233, 223, 11100, 5562, 1881, 245, 2129, 234, 32681, 38846, 34719, 6723, 41509, 27505, 8802, 10928, 24726, 4307, 121, 33597, 13876, 1543, 51109], "temperature": 0, "avg_logprob": -0.08952753360454853, "compression_ratio": 1.147239263803681, "no_speech_prob": 1.6501086955167388e-11}, {"id": 39, "seek": 14716, "start": 162.04, "end": 166.0, "text": "输出中间还有一段很值得看画面没有明显变化时,", "tokens": [51109, 9830, 241, 7781, 5975, 31685, 14852, 32241, 28427, 4563, 40242, 5916, 4200, 27126, 8833, 17944, 11100, 1431, 122, 2129, 246, 23756, 15729, 11, 51307], "temperature": 0, "avg_logprob": -0.08952753360454853, "compression_ratio": 1.147239263803681, "no_speech_prob": 1.6501086955167388e-11}, {"id": 40, "seek": 14716, "start": 166.0, "end": 170.44, "text": "终端没有继续增加描述,后面的人物进入以后文字才重新出现。", "tokens": [51307, 10115, 230, 11957, 107, 17944, 10115, 100, 10115, 255, 24228, 252, 9990, 11673, 237, 3316, 108, 11, 13547, 8833, 29979, 23516, 36700, 14028, 3588, 13547, 17174, 22381, 18888, 12624, 12560, 7781, 20204, 1543, 51529], "temperature": 0, "avg_logprob": -0.08952753360454853, "compression_ratio": 1.147239263803681, "no_speech_prob": 1.6501086955167388e-11}, {"id": 41, "seek": 17044, "start": 170.44, "end": 175.56, "text": "原始接口会把这种状态写成控制标记,真正做成应用时需要先过滤,", "tokens": [50365, 19683, 15476, 14468, 18144, 12949, 16075, 5562, 39810, 35276, 114, 3757, 223, 5676, 247, 11336, 48707, 25491, 162, 3921, 34756, 11, 6303, 15789, 10907, 11336, 44297, 9254, 15729, 35748, 10108, 16866, 15868, 97, 11, 50621], "temperature": 0, "avg_logprob": -0.11605156929262223, "compression_ratio": 1.2960526315789473, "no_speech_prob": 1.712588404423343e-11}, {"id": 42, "seek": 17044, "start": 175.56, "end": 178.76, "text": "对观众来说更重要的是模型知道什么时候先观察,", "tokens": [50621, 8713, 42350, 7384, 245, 6912, 8090, 19002, 24928, 24620, 41908, 39823, 7758, 10440, 29111, 10108, 42350, 47550, 11, 50781], "temperature": 0, "avg_logprob": -0.11605156929262223, "compression_ratio": 1.2960526315789473, "no_speech_prob": 1.712588404423343e-11}, {"id": 43, "seek": 17044, "start": 178.76, "end": 180.68, "text": "而不是每一帧都重复说话。", "tokens": [50781, 11070, 7296, 23664, 2257, 4845, 100, 7182, 12624, 1787, 235, 8090, 21596, 1543, 50877], "temperature": 0, "avg_logprob": -0.11605156929262223, "compression_ratio": 1.2960526315789473, "no_speech_prob": 1.712588404423343e-11}, {"id": 44, "seek": 17044, "start": 180.68, "end": 184.68, "text": "不过,实时不代表每个瞬间都不会漏掉,当前测试每秒只取一帧,", "tokens": [50877, 1960, 16866, 11, 24726, 15729, 1960, 19105, 17571, 23664, 7549, 36733, 105, 31685, 7182, 47928, 14065, 237, 29327, 11, 16233, 8945, 11038, 233, 5233, 243, 23664, 20956, 14003, 29436, 2257, 4845, 100, 11, 51077], "temperature": 0, "avg_logprob": -0.11605156929262223, "compression_ratio": 1.2960526315789473, "no_speech_prob": 1.712588404423343e-11}, {"id": 45, "seek": 17044, "start": 184.68, "end": 187.56, "text": "视频里一个非常短的击打动作就没有被稳定识别。", "tokens": [51077, 40656, 39752, 15759, 20182, 14392, 5881, 255, 1546, 6336, 119, 12467, 34961, 11914, 3111, 17944, 23238, 10415, 111, 12088, 5233, 228, 18453, 1543, 51221], "temperature": 0, "avg_logprob": -0.11605156929262223, "compression_ratio": 1.2960526315789473, "no_speech_prob": 1.712588404423343e-11}, {"id": 46, "seek": 17044, "start": 187.56, "end": 191.56, "text": "它看到了动作前后不等于看见了中间的每一刻,", "tokens": [51221, 11284, 18032, 2289, 34961, 11914, 8945, 13547, 1960, 10187, 37732, 4200, 23813, 2289, 5975, 31685, 1546, 23664, 2257, 45500, 11, 51421], "temperature": 0, "avg_logprob": -0.11605156929262223, "compression_ratio": 1.2960526315789473, "no_speech_prob": 1.712588404423343e-11}, {"id": 47, "seek": 19156, "start": 191.56, "end": 196.68, "text": "提高采样频率可能保留更多细节,也会同时增加输入和缓存压力。", "tokens": [50365, 20949, 12979, 5873, 229, 14496, 39752, 44866, 16657, 24302, 24456, 19002, 6392, 10115, 228, 45161, 11, 6404, 12949, 13089, 15729, 24228, 252, 9990, 9830, 241, 14028, 12565, 38109, 241, 39781, 5014, 233, 13486, 1543, 50621], "temperature": 0, "avg_logprob": -0.06883793366261018, "compression_ratio": 1.2944297082228118, "no_speech_prob": 1.6910110464674055e-11}, {"id": 48, "seek": 19156, "start": 196.68, "end": 202.76, "text": "这种持续更新来自模型的输入方式,视觉信息不是等到整段结束再交给语言部分,", "tokens": [50621, 5562, 39810, 17694, 10115, 255, 19002, 12560, 6912, 9722, 41908, 39823, 1546, 9830, 241, 14028, 9249, 27584, 11, 40656, 24447, 17665, 26460, 7296, 10187, 4511, 27662, 28427, 45641, 44320, 8623, 28455, 23197, 5233, 255, 12009, 32174, 11, 50925], "temperature": 0, "avg_logprob": -0.06883793366261018, "compression_ratio": 1.2944297082228118, "no_speech_prob": 1.6910110464674055e-11}, {"id": 49, "seek": 19156, "start": 202.76, "end": 208.68, "text": "每个采样画面还带着自己的时间位置,所以事件先后和状态变化能够进入同一段推理。", "tokens": [50925, 23664, 7549, 5873, 229, 14496, 27126, 8833, 14852, 4845, 99, 20708, 17645, 1546, 44848, 11160, 34719, 11, 7239, 6973, 20485, 10108, 13547, 12565, 35276, 114, 3757, 223, 2129, 246, 23756, 8225, 1787, 253, 36700, 14028, 13089, 2257, 28427, 33597, 13876, 1543, 51221], "temperature": 0, "avg_logprob": -0.06883793366261018, "compression_ratio": 1.2944297082228118, "no_speech_prob": 1.6910110464674055e-11}, {"id": 50, "seek": 19156, "start": 208.68, "end": 213.56, "text": "官方代码除了读取本地视频还支持摄像头、屏幕和外部发送的视频帧,", "tokens": [51221, 31929, 9249, 19105, 23230, 223, 32999, 2289, 5233, 119, 29436, 8802, 10928, 40656, 39752, 14852, 35488, 34783, 226, 12760, 39862, 1231, 9636, 237, 40211, 12565, 12022, 13470, 28926, 29309, 1546, 40656, 39752, 4845, 100, 11, 51465], "temperature": 0, "avg_logprob": -0.06883793366261018, "compression_ratio": 1.2944297082228118, "no_speech_prob": 1.6910110464674055e-11}, {"id": 51, "seek": 19156, "start": 213.56, "end": 219.0, "text": "它可以成为直播辅助实验观察、流程提醒或者屏幕任务监控的视觉底座。", "tokens": [51465, 11284, 6723, 11336, 13992, 16186, 49993, 9830, 227, 37618, 24726, 49657, 234, 42350, 47550, 1231, 27854, 29649, 20949, 40796, 31148, 9636, 237, 40211, 26443, 5087, 94, 5419, 239, 48707, 1546, 40656, 24447, 22816, 6346, 100, 1543, 51737], "temperature": 0, "avg_logprob": -0.06883793366261018, "compression_ratio": 1.2944297082228118, "no_speech_prob": 1.6910110464674055e-11}, {"id": 52, "seek": 21900, "start": 219.0, "end": 225.96, "text": "模型负责持续理解和回答,真正上线是告紧规则、权限控制和人工复合仍然要由应用层补上。", "tokens": [50365, 41908, 39823, 18464, 253, 18464, 96, 17694, 10115, 255, 13876, 17278, 12565, 8350, 28223, 11, 6303, 15789, 5708, 16853, 123, 1541, 16846, 7732, 100, 6758, 226, 2437, 247, 1231, 4422, 225, 43446, 48707, 25491, 12565, 4035, 23323, 1787, 235, 14245, 1550, 235, 5823, 4275, 23786, 44297, 9254, 9636, 224, 9890, 98, 5708, 1543, 50713], "temperature": 0, "avg_logprob": -0.09523574120359994, "compression_ratio": 1.1909814323607426, "no_speech_prob": 1.4807471221400803e-11}, {"id": 53, "seek": 21900, "start": 225.96, "end": 232.04, "text": "所以Mousewell Real-Time的优势很明确,视频还在发声,问题和答案就能一起向前推进。", "tokens": [50713, 7239, 44, 1316, 6326, 8467, 12, 22233, 1546, 7384, 246, 5087, 123, 4563, 11100, 38114, 106, 11, 40656, 39752, 14852, 3581, 28926, 32045, 11, 34069, 12565, 28223, 28899, 3111, 8225, 29567, 24282, 8945, 33597, 36700, 1543, 51017], "temperature": 0, "avg_logprob": -0.09523574120359994, "compression_ratio": 1.1909814323607426, "no_speech_prob": 1.4807471221400803e-11}, {"id": 54, "seek": 21900, "start": 232.04, "end": 240.04, "text": "它的限制也同样明确,采样会漏掉短动作,长上下文会继续消耗缓存,原始输出也不是装好就能使用的监控产品。", "tokens": [51017, 45224, 43446, 25491, 6404, 13089, 14496, 11100, 38114, 106, 11, 5873, 229, 14496, 12949, 14065, 237, 29327, 5881, 255, 34961, 11914, 11, 32271, 5708, 4438, 17174, 12949, 10115, 100, 10115, 255, 28837, 4450, 245, 38109, 241, 39781, 11, 19683, 15476, 9830, 241, 7781, 6404, 7296, 8083, 227, 2131, 3111, 8225, 22982, 9254, 1546, 5419, 239, 48707, 1369, 100, 30246, 1543, 51417], "temperature": 0, "avg_logprob": -0.09523574120359994, "compression_ratio": 1.1909814323607426, "no_speech_prob": 1.4807471221400803e-11}, {"id": 55, "seek": 21900, "start": 240.04, "end": 245.48, "text": "如果你的任务只是等视频结束以后再总结,指令版会更直接。", "tokens": [51417, 13119, 18961, 26443, 5087, 94, 36859, 10187, 40656, 39752, 45641, 44320, 3588, 13547, 8623, 33440, 45641, 11, 25922, 49061, 42096, 12949, 19002, 43297, 1543, 51689], "temperature": 0, "avg_logprob": -0.09523574120359994, "compression_ratio": 1.1909814323607426, "no_speech_prob": 1.4807471221400803e-11}, {"id": 56, "seek": 24548, "start": 245.48, "end": 252.44, "text": "如果你需要在视频进行中观察、提问等待,再根据新画面更新判断,实施版才真正对口。", "tokens": [50365, 45669, 35748, 3581, 40656, 39752, 36700, 8082, 5975, 42350, 47550, 1231, 20949, 22064, 10187, 18390, 11, 8623, 31337, 26075, 106, 12560, 27126, 8833, 19002, 12560, 2437, 97, 4307, 255, 11, 24726, 4307, 121, 42096, 18888, 6303, 15789, 8713, 18144, 1543, 50713], "temperature": 0, "avg_logprob": -0.10718824646689674, "compression_ratio": 0.9416666666666667, "no_speech_prob": 1.662408405378457e-11}, {"id": 57, "seek": 25244, "start": 252.44, "end": 257.4, "text": "项目地址和官方资料放在评论区,感兴趣可以拿自己的视频试一遍。", "tokens": [50365, 10178, 117, 11386, 10928, 14872, 222, 12565, 31929, 9249, 5266, 226, 33404, 12744, 3581, 5233, 226, 7422, 118, 9937, 118, 11, 9709, 2347, 112, 39835, 6723, 24351, 17645, 1546, 40656, 39752, 5233, 243, 2257, 3330, 235, 1543, 50613], "temperature": 0, "avg_logprob": -0.10118260383605956, "compression_ratio": 0.8888888888888888, "no_speech_prob": 1.526043180710701e-11}], "language": "zh"}