A $6.3 billion open-weight model just got embarrassed by the French...
FFireship@Fireship · 4.3M subscribers
- Views
- 1.5M
- 1,536,981
- Outlier
- ×1.5
- Median 1M
- Length
- 6:10
- Published
- Oct 7
- Wednesday
- Comments
- 906
- 20.9K likes
Script
How it's written
- Words
- 1,268
- Words a minute
- 205
- Reading time
- 6 min
- Segments
- 4
The hook
0:00–0:26Yesterday, while you were being force-fed every iteration of that Rumpelstiltskin video, the Paris-based AI lab Mistral released Mistral large four, a trillion-parameter open-weight model that they're claiming is the best of its kind, assuming you adopt the American ideology of just ignoring any technology coming out of China. It's natively multimodal, has a million-token context window, and claims to blah blah blah. The more interesting part of this release was the timing because it landed
Structure
Yesterday, while you were being force-fed every iteration of that Rumpelstiltskin video, the Paris-based AI lab Mistral released Mistral large four, a trillion-parameter open-weight model that they're claiming is the best of its kind, assuming you adopt the American ideology of just ignoring any technology coming out of China. It's natively multimodal, has a million-token context window, and claims to blah blah blah. The more interesting part of this release was the timing because it landed in the middle of what many are calling the dumbest week in the history of the open-weight race. It kicked off early in the week when Trump announced a superintelligence force whose job it is to keep America at the frontier of the AI race, with their first order of business claiming that AI is actually dumb and the real future is SI. Then Reflection AI, the $25 billion Nvidia-backed startup that calls itself America's open frontier lab, finally shipped their first model after 2 years of fundraising. It's called Beam. It has 501 billion parameters. It was trained on 6.3 billion dollars worth of GPUs. It benchmarked itself against an already outdated Chinese model. And 24 hours later, it got out-benchmarked itself by Mistral. Then, just yesterday, Moonshot AI, the Chinese lab behind Kimi, closed its final funding round at a $50 billion valuation and started prepping a Hong Kong IPO, which I'm told the superintelligence force has no interest investing in. In today's video, we'll look at what the French, the Americans, and the Chinese each shipped this week, find out why every government on Earth suddenly wants an open model to call its
Watch from 0:00own, and figure out which one you should actually use. It is October 7th, 2026, and you're watching The Code Report. If you're still not convinced about the benefits of open-weight models, last month a man in San Francisco learned them the hard way when he had the authorities looking for him after a joking conversation he had with Claude triggered a safety review. And just last week, a woman in Florida got arrested just days after telling Claude she was going to do something that's typically frowned upon in modern society. And even if you're not planning on committing a felony with proprietary models, you're still sending every prompt to a data center, which is scanned by a classifier, but potentially read by a safety team, and then used in whatever way the lab wants, like training future models, or kicking off the process to throw you in jail. And whole ass countries are having the same problem, too. If take a government that wants to use AI to find issues in their own power grid, using a model like Mythos, which is designed for this exact purpose, would require handing a foreign company your private schematics, and then hoping that they never get leaked, or that someone on their payroll isn't a foreign spy, which Dario himself has said may already be the case. Where if you were using an open weight model, most of these problems go away, which is why nearly every country in the world is trying to build their own if they can, or find ones they can trust if they can't. Which brings us back to this week and the great model mid-off. The first announcement came from Reflection AI, which was started by two DeepMind researchers who raised 4.6 billion, billion with a B, on the premise that the best open model in the world should come from the West, and Llama 4 wasn't cutting it. Well, this week we finally got to witness manifest destiny realized
Watch from 1:31when they released Beam. It's got 501 billion parameters, and its base model was trained in just under 4 weeks on about 24 trillion tokens. After that, they spent 4 more weeks post training it to get it better at writing code, using tools, and finishing long agent tasks. The result is a model that Reflection says keeps up with Jipu AI's GLM 5.2, while needing only a quarter of the hardware. But sadly, as of today, you still need an invite to use it, and the weights aren't actually out yet. Which means the best open American model is somehow still Gemma 4, which Google released under Apache 2.0 back in April. Then the next day, Mistral showed up with something that sounds like you'd order it from McDonald's. It's a mixture of experts model, which means it has about a trillion parameters, but only 49 billion of them actually turn on for a given token. So, the idea is that you get a trillion parameters worth of knowledge for roughly the cost of running a model 20 times smaller, which if that sounds familiar, it's probably because it's the same trick DeepSeek used last year to tank Nvidia. On the TMBB's, it barely beats GLM 5.3 at long coding jobs. It beats Beam by 18 points and lands in the top five for finding security bugs, which is the one Mistral actually cares about because the power grid scenario from earlier is the thing they're actually trying to sell. The catch is the numbers really are, trust me, bros, because they were provided by Mistral and they actually haven't finished training yet and the weights don't come out until later this month under their own custom license, which brings us to China, which already has the thing both of them are trying to build with Kimiko 3. It's a 2.8 trillion
Watch from 3:02parameter open model that Moonshot released back in July, which became so popular they had to stop taking new subscribers 2 days later because they ran out of GPUs. Since then, Microsoft, Amazon, and Google have all been negotiating to host it. Anthropic accused them of running 300,000 requests through fake accounts to copy Claude and the company went from a $4 billion valuation last November to 50 billion yesterday with China's own state AI fund involved in the round. So, which one should you actually use, though? If all you've got is a gaming PC, go with Gemma 4 and get your money up. If you have your own data center and no compliance department, go with Kimiko 3. Or if you have a data center and a compliance department, then go with Mistral large 4. And if you're a patriot, Beam's weights are coming later this month. But no matter which model you choose, they'll need to get on the internet, which they can do with Kernel, the sponsor of today's video. They provide open-source browser infrastructure that lets your agents act on the internet blazingly fast, but none of that actually matters if your agent keeps getting blocked by bot checks, which is why they just launched config registry. Send it the URL you want to access from either their dashboard or the API and it'll give you back the browser and proxy settings that have worked for that site. Right now, my agent keeps getting stopped by a bot check, but when I spin up a browser with the config Kernel recommends, it's able to access the page without any issues. The kernel builds these configs by continuously testing real browser and proxy setups against the live site, and each one comes with a success rate, a sample size, and last verified date. Use the code Fireship at the link below to get $30 in free credits, which covers your first month on the hobbyist plan. This has been the Code Report. Thanks for watching, and I will see you in the next one.
Watch from 4:32Full transcript
- 0:00
Yesterday, while you were being force-fed every iteration of that Rumpelstiltskin video, the Paris-based AI lab Mistral released Mistral large four, a trillion-parameter open-weight model that they're claiming is the best of its kind, assuming you adopt the American ideology of just ignoring any technology coming out of China. It's natively multimodal, has a million-token context window, and claims to blah blah
- 0:22
blah. The more interesting part of this release was the timing because it landed in the middle of what many are calling the dumbest week in the history of the open-weight race. It kicked off early in the week when Trump announced a superintelligence force whose job it is to keep America at the frontier of the AI race, with their first order of business claiming that AI
- 0:42
is actually dumb and the real future is SI. Then Reflection AI, the $25 billion Nvidia-backed startup that calls itself America's open frontier lab, finally shipped their first model after 2 years of fundraising. It's called Beam. It has 501 billion parameters. It was trained on 6.3 billion dollars worth of GPUs. It benchmarked itself against an already
- 1:04
outdated Chinese model. And 24 hours later, it got out-benchmarked itself by Mistral. Then, just yesterday, Moonshot AI, the Chinese lab behind Kimi, closed its final funding round at a $50 billion valuation and started prepping a Hong Kong IPO, which I'm told the superintelligence force has no interest investing in. In today's video, we'll look at what the French, the Americans,
- 1:25
and the Chinese each shipped this week, find out why every government on Earth suddenly wants an open model to call its own, and figure out which one you should actually use. It is October 7th, 2026, and you're watching The Code Report. If you're still not convinced about the benefits of open-weight models, last month a man in San Francisco learned them the hard way when he had the authorities looking for him after a joking conversation he had with Claude
- 1:47
triggered a safety review. And just last week, a woman in Florida got arrested just days after telling Claude she was going to do something that's typically frowned upon in modern society. And even if you're not planning on committing a felony with proprietary models, you're still sending every prompt to a data center, which is scanned by a classifier, but potentially read by a safety team, and then used in whatever
- 2:07
way the lab wants, like training future models, or kicking off the process to throw you in jail. And whole ass countries are having the same problem, too. If take a government that wants to use AI to find issues in their own power grid, using a model like Mythos, which is designed for this exact purpose, would require handing a foreign company your private schematics, and then hoping that they never get leaked, or that
- 2:27
someone on their payroll isn't a foreign spy, which Dario himself has said may already be the case. Where if you were using an open weight model, most of these problems go away, which is why nearly every country in the world is trying to build their own if they can, or find ones they can trust if they can't. Which brings us back to this week and the great model mid-off. The first announcement came from Reflection AI,
- 2:47
which was started by two DeepMind researchers who raised 4.6 billion, billion with a B, on the premise that the best open model in the world should come from the West, and Llama 4 wasn't cutting it. Well, this week we finally got to witness manifest destiny realized when they released Beam. It's got 501 billion parameters, and its base model was trained in just under 4 weeks on
- 3:08
about 24 trillion tokens. After that, they spent 4 more weeks post training it to get it better at writing code, using tools, and finishing long agent tasks. The result is a model that Reflection says keeps up with Jipu AI's GLM 5.2, while needing only a quarter of the hardware. But sadly, as of today, you still need an invite to use it, and the weights aren't actually out yet. Which
- 3:30
means the best open American model is somehow still Gemma 4, which Google released under Apache 2.0 back in April. Then the next day, Mistral showed up with something that sounds like you'd order it from McDonald's. It's a mixture of experts model, which means it has about a trillion parameters, but only 49 billion of them actually turn on for a given token. So, the idea is that you
- 3:50
get a trillion parameters worth of knowledge for roughly the cost of running a model 20 times smaller, which if that sounds familiar, it's probably because it's the same trick DeepSeek used last year to tank Nvidia. On the TMBB's, it barely beats GLM 5.3 at long coding jobs. It beats Beam by 18 points and lands in the top five for finding security bugs, which is the one Mistral
- 4:11
actually cares about because the power grid scenario from earlier is the thing they're actually trying to sell. The catch is the numbers really are, trust me, bros, because they were provided by Mistral and they actually haven't finished training yet and the weights don't come out until later this month under their own custom license, which brings us to China, which already has the thing both of them are trying to build with Kimiko 3. It's a 2.8 trillion
- 4:32
parameter open model that Moonshot released back in July, which became so popular they had to stop taking new subscribers 2 days later because they ran out of GPUs. Since then, Microsoft, Amazon, and Google have all been negotiating to host it. Anthropic accused them of running 300,000 requests through fake accounts to copy Claude and the company went from a $4 billion valuation last November to 50 billion
- 4:54
yesterday with China's own state AI fund involved in the round. So, which one should you actually use, though? If all you've got is a gaming PC, go with Gemma 4 and get your money up. If you have your own data center and no compliance department, go with Kimiko 3. Or if you have a data center and a compliance department, then go with Mistral large 4. And if you're a patriot, Beam's weights are coming later this month. But no matter which model you choose,
- 5:15
they'll need to get on the internet, which they can do with Kernel, the sponsor of today's video. They provide open-source browser infrastructure that lets your agents act on the internet blazingly fast, but none of that actually matters if your agent keeps getting blocked by bot checks, which is why they just launched config registry. Send it the URL you want to access from either their dashboard or the API and
- 5:36
it'll give you back the browser and proxy settings that have worked for that site. Right now, my agent keeps getting stopped by a bot check, but when I spin up a browser with the config Kernel recommends, it's able to access the page without any issues. The kernel builds these configs by continuously testing real browser and proxy setups against the live site, and each one comes with a success rate, a sample size, and last
- 5:57
verified date. Use the code Fireship at the link below to get $30 in free credits, which covers your first month on the hobbyist plan. This has been the Code Report. Thanks for watching, and I will see you in the next one.
Description
SponsorKERNEL is the simplest way to let your agents act on any website. Use code: FIRESHIP to get a free upgrade to the Hobbyist tier ($30 value) at https://kernel.sh/fireship
Mistral, Reflection AI, and Moonshot all made big moves in the open-weight race this week. Let's dive in.
Want more Fireship?
🗞️ Newsletter: https://bytes.dev
🧠 Courses: https://fireship.dev
More from Fireship
×4.27:27Did OpenAI actually build AGI? GPT-6 Astra first look
Fireship4.2M views
×4.15:56Every app on Earth is now open source... whether it likes it or not
Fireship4.1M views
×3.45:45OpenAI's biggest math breakthrough is getting ugly...
Fireship3.4M views
×3.37:22Tragic mistake... Anthropic leaks Claude’s source code
Fireship3.3M views
