WEBVTT 1 00:00:00.540 --> 00:00:01.710 So when we talk 2 00:00:01.710 --> 00:00:03.840 about open Large Language Models, 3 00:00:03.840 --> 00:00:05.940 we're talking about their weights 4 00:00:05.940 --> 00:00:09.930 or parameters being made publicly available. 5 00:00:09.930 --> 00:00:13.770 And, therefore, another important question, of course, is, 6 00:00:13.770 --> 00:00:17.790 which open models do exist out there, 7 00:00:17.790 --> 00:00:20.280 for which models have the weights, 8 00:00:20.280 --> 00:00:24.510 the parameters been made available publicly? 9 00:00:24.510 --> 00:00:26.670 Well, popular examples are, 10 00:00:26.670 --> 00:00:29.730 for example, Meta's Llama models. 11 00:00:29.730 --> 00:00:32.400 So Meta, Facebook previously, 12 00:00:32.400 --> 00:00:35.010 did create large language models, 13 00:00:35.010 --> 00:00:37.080 which they branded Llama. 14 00:00:37.080 --> 00:00:39.720 That's just the name of the model family 15 00:00:39.720 --> 00:00:42.600 that are indeed available publicly. 16 00:00:42.600 --> 00:00:45.510 Again, not the code that was used for training them, 17 00:00:45.510 --> 00:00:47.763 but the model weights and parameters. 18 00:00:48.810 --> 00:00:52.830 The same is true for Google with their Gemma models. 19 00:00:52.830 --> 00:00:56.460 Now, when thinking of Google and AI, 20 00:00:56.460 --> 00:00:59.310 Google Gemini might be the first thing 21 00:00:59.310 --> 00:01:01.890 that comes to your mind because Google Gemini 22 00:01:01.890 --> 00:01:04.740 is essentially their answer to ChatGPT. 23 00:01:04.740 --> 00:01:07.290 It's their AI chat bot. 24 00:01:07.290 --> 00:01:08.520 Or to be precise, 25 00:01:08.520 --> 00:01:11.880 Gemini, in general, is the brand Google uses 26 00:01:11.880 --> 00:01:15.360 for their proprietary AI models 27 00:01:15.360 --> 00:01:18.030 that are also used in that Gemini app. 28 00:01:18.030 --> 00:01:21.660 But that can, for example, also be accessed programmatically 29 00:01:21.660 --> 00:01:24.210 through the Gemini developer API. 30 00:01:24.210 --> 00:01:27.690 But Google does not just have Gemini. 31 00:01:27.690 --> 00:01:31.590 Gemini are their proprietary closed models 32 00:01:31.590 --> 00:01:33.630 where you don't get access to the weights. 33 00:01:33.630 --> 00:01:36.300 But they also have their Gemma models, 34 00:01:36.300 --> 00:01:38.670 which are open models 35 00:01:38.670 --> 00:01:42.000 where you can get the weights, the parameters, 36 00:01:42.000 --> 00:01:43.290 and where you can therefore 37 00:01:43.290 --> 00:01:46.320 run these models locally on your system. 38 00:01:46.320 --> 00:01:49.530 Another quite popular creator 39 00:01:49.530 --> 00:01:54.480 of a very popular open model is DeepSeek. 40 00:01:54.480 --> 00:01:57.570 Now, you may recall that in early 2025, 41 00:01:57.570 --> 00:02:00.150 everybody was talking about DeepSeek 42 00:02:00.150 --> 00:02:04.080 because this Chinese mysterious company 43 00:02:04.080 --> 00:02:07.590 released an open Large Language Model 44 00:02:07.590 --> 00:02:10.020 that all of a sudden rivaled 45 00:02:10.020 --> 00:02:13.020 the latest models published by OpenAI. 46 00:02:13.020 --> 00:02:16.350 And those latest models were not the open models, 47 00:02:16.350 --> 00:02:18.330 but their proprietary models. 48 00:02:18.330 --> 00:02:21.270 And all of a sudden DeepSeek came around the corner 49 00:02:21.270 --> 00:02:24.480 and published a rivaling state-of-the-art 50 00:02:24.480 --> 00:02:26.820 Large Language Model by making 51 00:02:26.820 --> 00:02:30.000 the parameters of that model publicly available. 52 00:02:30.000 --> 00:02:32.430 Something OpenAI and Google did not do 53 00:02:32.430 --> 00:02:35.760 with their state-of-the-art models back then. 54 00:02:35.760 --> 00:02:38.490 And whilst DeepSeek also published 55 00:02:38.490 --> 00:02:41.730 an online chat bot similar to ChatGTP, 56 00:02:41.730 --> 00:02:44.460 where they hosted the model for you, 57 00:02:44.460 --> 00:02:47.850 they did indeed also make the weights available publicly 58 00:02:47.850 --> 00:02:50.970 so that you could run it locally on your system, 59 00:02:50.970 --> 00:02:52.470 depending on your hardware, 60 00:02:52.470 --> 00:02:55.680 because R1 was quite a large model. 61 00:02:55.680 --> 00:02:58.560 But I'll get back to the hardware requirements later. 62 00:02:58.560 --> 00:03:02.220 And you can indeed run many of these open models 63 00:03:02.220 --> 00:03:03.780 on your laptop as well. 64 00:03:03.780 --> 00:03:06.270 I can already say as much. 65 00:03:06.270 --> 00:03:08.910 But DeepSeek, therefore, is another important name 66 00:03:08.910 --> 00:03:13.170 to remember when talking about open Large Language Models. 67 00:03:13.170 --> 00:03:17.040 Another quite popular name would be Mistral, 68 00:03:17.040 --> 00:03:19.560 which is a European AI company 69 00:03:19.560 --> 00:03:22.920 that also has some proprietary closed models, 70 00:03:22.920 --> 00:03:27.090 but also some publicly available open models. 71 00:03:27.090 --> 00:03:30.960 And, of course, there are many, many more. 72 00:03:30.960 --> 00:03:33.360 And the question therefore, of course, is, 73 00:03:33.360 --> 00:03:36.240 where do you find those open models? 74 00:03:36.240 --> 00:03:38.610 And then, of course, also how do you use them? 75 00:03:38.610 --> 00:03:41.433 But let's start with the finding part.