The open model war between Europe, the US and China heats up

French AI company Mistral unveiled Mistral Large 4 on October 6. Its nickname is “Le Chonk”: a 1-trillion-parameter model that also understands images. Mistral calls it “the best open weights model from US or Europe.”

It was trained end to end in Europe, in Mistral’s own data centers.

For now, though, it’s only a preview: the model is available through the API, and the weights come at the end of October. It’s a mixture-of-experts model, so only about 5% of its 1 trillion parameters run at each step.

Even Mistral’s own chart is mixed: on the DeepSWE coding test, Large 4 scores 62 while China’s Kimi K3 leads with 68.

On independent benchmarker Artificial Analysis’s Intelligence Index, Large 4 scores 38, making it the smartest model built outside the US and China. Its predecessor, Mistral Large 3, scored 9.

But seven Chinese models rank above it on the same list, including GLM-5.3, Qwen3.8 Max and Kimi K3. It also costs more than 4x as much per task as open models of similar intelligence.

The same day, news came from the opposite direction. Christian Kroll, founder and CEO of German search engine Ecosia, told Politico: “We are disappointed with the quality of Mistral.” In his view, Mistral’s models are “a year behind” the competition.

Ecosia is moving its AI features to open models, including Chinese ones, through German platform Melious, which hosts Qwen, GLM and Kimi.

Quality wasn’t the only reason. Kroll says Mistral’s servers were often overloaded: “We were simply too large a customer for Mistral.” With the new setup, he says, Ecosia roughly halved its costs while improving quality. He also doesn’t see Mistral as “truly sovereign,” given its international investors.

Ecosia only switched from OpenAI to Mistral in May. Mistral’s Guillaume Lample encouraged Ecosia to test the new model as soon as possible.

A move came from across the Atlantic too. Nvidia-backed US startup Reflection AI unveiled its open model Beam on October 5: 501 billion parameters, 23 billion active at each step. The weights are due this month under an Apache 2.0 license.

The company says Beam “advances the Western open frontier.” Semafor describes Reflection as “America’s answer to open-source Chinese AI.”

But Beam doesn’t beat the Chinese models everywhere. In Reflection’s own table, Kimi K3 leads on almost every test, and the company says so openly. Beam’s pitch isn’t raw power, it’s efficiency: results comparable to GLM 5.2 with 3 to 4 times less compute.

Artificial Analysis says early signs point to Beam being one of the most token-efficient open models for its level of intelligence. Independent testing is ongoing.

China isn’t standing still either. According to Bloomberg, DeepSeek is close to raising at least 80 billion yuan (about $12 billion), with Tencent and CATL among the biggest backers. It started the round in July targeting a valuation of about $74 billion and plans an IPO in early 2027. DeepSeek hasn’t commented.