Reflection Beam and Cohere North 2: Open Weights vs Locked Down AI

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Reflection Beam and Cohere North 2: Open Weights vs Locked Down AI · Channel: Fahd Mirza · 09:30 · Watch on YouTube

Topics: Enterprise AI applications, AI models, Local and open-source models, AI safety

Summary

The video compares Cohere North 2, an enterprise-focused platform built around security and control, with Reflection AI’s announced Beam model, whose weights are promised as an open release later this month. North 2 emphasizes deployment options, governance tools, predictable costs, and the ability to use different models. Beam is presented as a 501-billion-parameter mixture-of-experts model trained with extensive pre-training and reinforcement learning, though its weights are not yet available and its benchmark results do not lead every comparison. The video argues that both releases respond to concerns about dependence on a small number of closed AI providers, and says Beam’s claims need to be checked through independent testing.

Key points

Cohere North 2 is an agentic platform that can run in the cloud, on premise, or in an air-gapped environment.
North 2 offers reusable agents, memory, skills, integrations with workplace tools, spending controls, and support for bringing a model of choice.
Reflection AI says Beam has 501 billion total parameters, with 23 billion active at a time, using a mixture-of-experts design.
Reflection says Beam’s weights will be released later this month under the Apache 2 license; at the time of the video, they were not available.
The company reports pre-training on 23.8 trillion tokens and a reinforcement-learning run involving 10.5 thousand GPUs and over 100 million attempts.
The presenter highlights Reflection’s claims about data filtering, a smooth training-loss curve, and time spent on useful training.
Beam leads on many tests in Reflection’s comparisons, but the presenter notes that some larger open models perform better on certain benchmarks and capabilities.
The video frames both releases as responses to concerns about being locked into a small number of closed AI providers.
The presenter says Beam should be judged by whether its weights arrive as promised and whether independent tests confirm its results.

Chapters

Start	Chapter	Summary
00:01	Two releases, one industry pressure	The video introduces Cohere’s focus on enterprise control and Reflection AI’s promised open model, against a backdrop of pressure to make AI businesses profitable.
01:35	Cohere North 2 and enterprise control	The presenter describes North 2 as a flexible agentic platform offering deployment control, integrations, governance features, and predictable spending.
03:13	Reflection AI announces Beam	The video outlines Beam’s promised Apache 2 weight release, its mixture-of-experts design, and Reflection’s claims about its training data and infrastructure.
05:47	Reinforcement learning and benchmark results	The presenter reviews Reflection’s reported reinforcement-learning run and benchmark claims, while noting that Beam does not outperform every larger open model.
07:54	The open-weights test	The video connects the two releases as responses to provider lock-in and says Beam’s weights and independent test results will determine whether its promise holds up.

Quotes

No more trade-offs between security, intelligence, cost, and control. — 02:08

So, treat this as a promise until the files are actually downloadable. — 03:13

The message they are trying to give is that data quality matters more than the data volume — 04:45

Both are reactions to the same fear, by the way. — 08:25

Will the weights actually land this month and will the results hold up when independent people test them? — 08:58

Concepts and entities

Cohere, North 2, Reflection AI, Beam, Apache License 2.0, Mixture of Experts, reinforcement learning, Mistral, NVIDIA

Full transcript (original language)

00:01 Two AI releases landed today and they both tell the same story from opposite ends. One is from Cohere and it is all about control, governance, and keeping your data locked inside your walls. The other is from Reflection AI and it is a 501 billion parameter model that they say will be fully open. Why does this matter right now? Because the AI industry is under extreme

00:32 pressure to prove it can make money. Enterprise AI today is dominated by two companies. OpenAI and Anthropic. Both are spending enormous amount of money and many reports say they’re still running at heavy losses. The investors who funded all of this now want their money back or at least returns. So, what do companies do under that pressure? Many keep their best models closed. Some delay releasing weights and

01:04 you can see that with almost all of these Chinese companies who are now very reluctant to open source and open with their models or at least their first go with their token plans for a month or so and only then they release the models. Other companies are pivoting hard towards enterprise, especially the European and non-American companies. Excluding OpenAI and Anthropic which were always enterprise plus the large-scale user base.

01:35 These companies are selling control, security, and cost dashboard because that is where the budgets are and one glaring example is Mistral. That is the backdrop for today. Let’s start with the enterprise side. This is Fahad Mirza and I welcome you to the channel. If you want to help out the channel, please become a member and please subscribe to this free weekly AI newsletter. So, first up, let’s talk about this North 2 from Cohere. Cohere launched

02:08 North 2 and it is a perfect example of the enterprise pivot. It is not a new model. It is a full agentic platform that you can run in the cloud, on premise, or fully air-gapped, which means completely cut off from the internet. And I can tell you government agencies, at least here in Australia, would love such model. The pitch is very simple. No more trade-offs between security, intelligence, cost, and control. You get

02:41 reusable agents, memory, skills, connectors to you know, tools like Teams, Slack, SharePoint, GitHub, and admin panel where you can set token spending limits per user. This, you know, North 2 model is also, you know, the platform I mean, is geared towards becoming model agnostic. So, you can even bring your own model. So, notice what they are selling, not raw intel intelligence, governance, and

03:13 predictable bills that tells you where the money in AI is heading. Now, let’s come back to the main event which I wanted to talk about. Reflection AI has just announced Beam, their first open weight model. As of today, the weights are not out. Reflection says they will release them later this month under an Apache 2 license, which is very permissive and always very welcome. So, treat this as a

03:43 promise until the files are actually downloadable. Everything I quote next comes from Reflection’s own announcement, not independent in setting. So, what exactly is this model? Be has 501 billion total parameters, but only 23 billion are active at any moment. It uses a design called as mixture of expert, which we already know. Um, you know, for those who really don’t

04:13 know and don’t have PhD in machine learning, one example I really love about mixture of expert is that it is sort of a hospital with hundreds of specialist doctors. For each patient, only a few specialists that called in. You get the knowledge of the whole hospital, but pay for only a few doctors at a time. That is why a huge model can still be cheap to run locally or even through the API. Now, this Be

04:45 was pre-trained on 23.8 trillion tokens on a cluster of 6,144 Nvidia GP300 GPUs in under 4 weeks. That’s it, 4 weeks. They say about 95% of raw internet data was thrown away during cleaning, and that standard filters would have missed roughly 1.8 trillion good tokens they kept. The message they are trying to give is that data quality matters more than the

05:15 data volume, and I wholeheartedly agree. Look at the pre-training loss um on your screen. Loss is basically how wrong the model is, and lower is better. Big training runs usually have very ugly spikes where things nearly blow up. This curve is quite smooth. They report only nine restarts in the whole run, and 92.3% of the time spent on useful training. That is serious engineering.

05:47 Now, one thing they have talked a lot about is reinforcement learning and I think that is the heart of Beam. After pre-training, they taught the model through trial and error, which is called reinforcement learning. Imagine, um or just, you know, think about a student who solves a million practice problems and gets graded on each one. Reflection ran this on 10.5 thousand GPUs for 4 weeks generating over 100 million attempts using about 1

06:18 million different practice environments. They claim it is one of the largest other RL runs by any open lab and that scores were still climbing when they stopped. This is quite impressive. And here comes a very interesting number screen. These are reflections on published results against the other open models in their table like Inkling, Nemotron-3, and Draw from Nvidia. Beam leads on most tests.

06:49 Like you can see Sweep Bench, amazing. This is 80.9, Terminal Bench 2.1 is 80.1, and so on and so forth. But, you know, the thing is that um what I have found, especially in reasoning, tool call, and general capability, and also a few others, and Reflection admits it, too, against the bigger open models, Beam does not win. On Terminal Bench, even GLM-5.2 scores 81.2, 81.0, Qwen-3.8-Max scores 86.6, and

07:22 Kimiko-3 is really throwing it out of the park at 88.3, which is quite impressive. Another thing, uh which they also have said, is that um Kimiko-3 is ahead on raw capability. Beam’s argument is not that it is it is that it is the most efficient and the best Western open weight option. We will see when it gets released. We we are we we will install it locally if it is possible. If

07:54 not, we will at least thoroughly test it out and you will see that video. Now look, even more than both of these model releases, my point of view is slightly different as how this these two releases fit together. I normally don’t combine two models in one video, but I wanted to for this one. First up, Cohere says that your AI should run inside your walls, under your rules, within your budget controls.

08:25 Reflection says the model itself should be open, so anyone can run it anywhere, but they have delayed the open weights. I don’t know why, maybe they’re still training or they just wanted to have some API stuff or something like that beforehand. Both are reactions to the same fear, by the way. Which is being locked into a small number of closed providers while the economics of AI are still unproven. And that is a key thing. The test for Beam is simple.

08:58 Will the weights actually land this month and will the results hold up when independent people test them? Like us. If yes, that is a big moment for OpenAI. I have no doubt about it. Especially looking at these numbers. Another thing is that if that doesn’t happen, it becomes one more promise in an industry full of them. Especially you can see the pivot of Cohere. And if you don’t know, we have we have covered Cohere a lot.

09:30 They started with sort of multilingual models and it was quite good, especially in non-English language languages, but now they are pivoting towards this sort of thing. We’ll be watching closely, and if you want me to test Beam hands-on on the day the weights drop, tell me in the comments, become a member to support the channel. Thank you for all the support.