TECH & SCIENCE
PARIS / NEW YORK
It has a meme for a nickname, a trillion parameters for a brain and a very serious job: convince the world that Western open AI can keep up with China. This week, Paris-based Mistral AI pulled the cover off Mistral Large 4 — affectionately and, according to its own launch post, "very officially" called le Chonk — and in the very same news cycle, Brooklyn startup Reflection AI revealed a rival called Beam.
Two launches in two days. Two companies promising powerful, cheaper-to-run models that businesses and governments can control themselves. And one shared target: the Chinese open-weight systems, from DeepSeek to Kimi to Qwen, that have dominated the free-to-download AI scene.
Welcome to the open-weight arms race, where the weapons are neural networks and the prize is who the world's companies and governments trust with their data.
Meet 'Le Chonk'
Mistral launched Large 4 on Tuesday, 6 October 2026. It uses a mixture-of-experts design with 1 trillion parameters in total, of which roughly 49 billion are active when answering any single query. That is a major step up from its predecessor, Mistral Large 3, which launched in December 2025 with 675 billion total and 41 billion active parameters. OfficeChai reports it is natively multimodal and available now through Mistral's API.
The weights — the files that let outsiders download and run the model themselves — are promised by the end of October. The licence has not been announced. Earlier Mistral models used Apache 2.0 or a modified MIT licence, so many developers will be watching closely.
Mistral's chief scientist, Guillaume Lample, wrote on X that Large 4 is "by far the strongest open-weight model from the US or Europe." Note the careful wording: the claim excludes Chinese labs, and Mistral's own launch materials mainly compare the model against Chinese open-weight rivals including DeepSeek V4 Pro, Kimi K3, GLM-5.3 and Qwen3.8 Max.
The meme that became a model
The name is the best part. In June 2026, after Mistral renamed its Le Chat assistant to Vibe, internet users invented a fictional model called "Le Chaton Fat" — roughly, the fat kitten. CEO Arthur Mensch played along, joking that the model was really called "le gros chaton," the big kitten. When the real thing arrived, Mistral leaned in, and its launch post lists the model, "very officially," as le Chonk.
It is a rare example of a billion-euro AI lab letting its community name its flagship — and a small sign of how personality is becoming part of the competition.
Price tag: the real headline
The cheeky name should not distract from the pricing, which is the part rivals will notice. According to figures reported at launch, Large 4 costs $1.36 per million input tokens and $4.18 per million output tokens. By comparison, Claude Opus 5.5 is listed at $4 for input and $20 for output, while GPT-6 Astra is listed at $10 and $50.
On those numbers, Large 4 runs at about a third of the input price of Claude Opus 5.5 and roughly a fifth of the output price. Against GPT-6 Astra, it is around a seventh on input and a twelfth on output. For a company running millions of AI requests a day, that is the difference between a rounding error and a budget crisis.
Cheap, of course, only matters if the model is good enough. So how does it stack up?
How good is it, really?
The early independent-style numbers, as reported, paint a picture of a strong contender rather than an outright champion. On Finance Agent v2, Large 4 scored 54.7, ahead of GPT-6 Astra at 53.5 but behind Claude Opus 5.5 at 58.6. In a blind human coding evaluation by Surge AI, it scored 3.74 out of 5 and ranked second of five, while Claude Opus 5 scored 4.22.
On AutomationBench, a test of 657 business-software tasks, Large 4 scored 59.9, trailing Claude Sonnet 5.5 at 71.8, Claude Opus 5.5 at 69.5 and Gemini 4 Argon at 77.5. On the DeepSWE 1.1 coding test it scored 62, ahead of GLM-5.3 at 61 and DeepSeek V4 Pro at 57, but behind Kimi K3 at 68. A separate leaderboard puts GPT-6 Astra and Claude Opus 5 at 74.
Mistral also says the model is state-of-the-art on "critical workloads" such as cyber defence, manufacturing and finance, and that it surpasses closed frontier models at visual grounding — finding and identifying objects in images. CEO Arthur Mensch told a conference in Abu Dhabi, as reported by Reuters, that the model beats Chinese models on cybersecurity. These are the company's own claims, and independent testing will take time.
The takeaway: Large 4 does not beat the best closed models across the board, but it is far cheaper and, if the weights land as promised, can be run on a company's own servers.
Enter Beam: Brooklyn's challenger
While Paris was still celebrating, Reflection AI announced Beam on 5 October. The two-year-old startup was founded in 2024 by two former Google DeepMind researchers and has raised roughly $4.7 billion, according to PitchBook, from backers including Nvidia, Sequoia Capital and Lightspeed Venture Partners. Its last round valued it at $25 billion pre-money, and it has signed compute deals worth more than $7 billion combined with SpaceX and Nebius for access to Nvidia GB300 chips through 2029.
Beam is its first frontier open-weight model. It is text-only, uses a mixture-of-experts design with 501 billion total parameters and 23 billion active, was pretrained on 23.8 trillion tokens, and offers a context window of 1 million tokens. It targets reasoning, coding and agentic tasks, and was trained with high-compute reinforcement learning.
Reflection says Beam matches Z.ai's GLM-5.2 on advanced reasoning benchmarks and beats leading Western open models, while using three to four times less inference compute than rivals. It also claims higher scores than Thinking Machines Lab's Inkling on four coding tests, though Inkling is multimodal and Beam is not. The South China Morning Post reported that Reflection calls Beam "competitive" with GLM-5.2 and "approaching" Alibaba's Qwen3.8-Max on coding and agentic work, while Moonshot AI's Kimi K3 remains ahead in raw capability. All of this comes from Reflection's own benchmarks and has not been independently verified.
Reflection says it will release Beam's weights and full technical details this month, distributing through hyperscalers and neoclouds with integrations across open-source libraries at launch.
'AI factories' and the sovereignty pitch
Both companies are selling more than software. They are selling independence. Mistral pitches "sovereign AI" that customers can run without sending data to outside firms, and says Large 4 was built end to end in Europe and can be deployed from Europe on Mistral Cloud. In August 2026 Saudi Arabia's state-backed HUMAIN signed a deal worth hundreds of millions of euros with the company, and in September Mistral raised a €3 billion Series D at a valuation above €21 billion, led by Samsung. Mistral says Large 4 is the first milestone funded by that round.
Reflection's version is "AI factories," in which institutions train its models on their own proprietary data to build customised, local systems. Its targets are enterprises and sovereign nations, and Axios reported interest from hedge funds and trading firms. Reflection is testing a sovereign AI factory partnership with Shinsegae Group in South Korea. Nvidia chief Jensen Huang, whose company backs Reflection, champions the AI factory concept — which, handily, also sells a lot of Nvidia chips.
The message to governments is blunt: do not depend on a foreign provider for something this important. It is a message that resonates in capitals from Riyadh to Seoul.
Why the West is racing
For the past couple of years, the strongest freely downloadable models have tended to come from Chinese labs. That has worried policymakers and executives in the US and Europe who would rather build on technology from their own side. Reflection explicitly positions itself against closed labs such as Anthropic and OpenAI, Chinese open-model developers such as DeepSeek, Qwen and Z.ai, and Western open-model players including Mistral, Meta and Cohere.
Mistral's launch also came days after Germany's Aleph Alpha released a model called Kolibri, according to OfficeChai, suggesting a broader European push. The competition is no longer just US versus China. It is a three-cornered scramble in which every side promises cheaper, more efficient and more controllable AI.
Expert perspective: read the fine print
Lample's boast that Large 4 is "by far the strongest open-weight model from the US or Europe" is bold, but it is also carefully bounded. It does not claim to beat Kimi K3, which scored higher on the coding test reported at launch. Reflection, for its part, concedes that Kimi K3 leads in raw capability.
That honesty is a clue to where the real battle lies. Open-weight models rarely win on peak intelligence; they win on cost, privacy and flexibility. A bank that cannot send customer data to a third party, or a government that wants to audit every line of its AI, may happily trade a few benchmark points for control.
Caution is also warranted because most numbers so far come from the companies themselves, and two key facts remain pending: Mistral's licence terms and the actual release of both sets of weights. Running a trillion-parameter model locally needs serious hardware, so for most developers the API will be the way in, at least at first.
What this means for you
If you use AI at work, the practical effect is downward pressure on prices. When a capable model costs a fraction of the leading closed options, rivals have to respond, and tools built on top of these systems can get cheaper or more powerful.
If you care about privacy, the rise of models that can run inside a company's or country's own infrastructure means your data may be less likely to travel to a foreign server. If you are a developer, October could bring two heavyweight open models to download, test and fine-tune.
And if you are simply watching the geopolitics, this is a reminder that AI leadership is not a settled question. A cat-meme model from Paris and a Nvidia-backed upstart from Brooklyn are now openly challenging the idea that China owns open AI.
The bottom line
The open-weight race just got a lot more crowded, and a lot more interesting. Mistral has the trillion-parameter headline, the low price and the meme. Reflection has the efficiency pitch, the billions in backing and the promise of AI factories.
Both now face the hard part: delivering the weights this month, surviving independent testing and proving that cheaper really can mean good enough. If they do, the big kitten and the Brooklyn challenger may change who builds the world's AI. If they don't, the Chinese labs will happily keep the crown.
